Skip to contentSkip to conversation

hw/block: m25p80: Fix the mess of dummy bytes needed for fast read commands

PATCH In review Bin Meng 7 participants 36 messages Jan 14, 2021 – Apr 28, 2021

Subsystems (2)
lore
Share & tools
lore

Bin Meng

author
#0 lore raw

cover hw/block: m25p80: Fix the mess of dummy bytes needed for fast read commands

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.
- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

This series fixes the mess in the m25p80 from the flash side first, followed by fixes to 3 known SPI controller models that fall into the 2nd case above.

Please note, I have no way to verify patch 7/8/9 because:

* There is no public datasheet available for the SoC / SPI controller
* There is no QEMU docs, or details that tell people how to boot either U-Boot or Linux kernel to verify the functionality

These 3 patches are very likely to be wrong. Hence I would like to ask help from the original author who wrote these SPI controller models to help testing, or completely rewrite these 3 patches to fix things. Thanks!

Patch 6 is unvalidated with QEMU, mainly because there is no doc to tell people how to boot anything to test. But I have some confidence based on my read of the ZynqMP manual, as well as some experimental testing on a real ZCU102 board.

Other flash patches can be tested with the SiFive SPI series:
http://patchwork.ozlabs.org/project/qemu-devel/list/?series=222391

Cherry-pick patch 16 and 17 from the series above, and switch to different flash model to test with the following command:

$ qemu-system-riscv64 -nographic -M sifive_u -m 2G -smp 5 -kernel u-boot

I've picked up two for testing:

QEMU flash: "sst25vf032b"

U-Boot 2020.10 (Jan 14 2021 - 21:55:59 +0800)

CPU: rv64imafdcsu
Model: SiFive HiFive Unleashed A00
DRAM: 2 GiB
MMC:
Loading Environment from SPIFlash... SF: Detected sst25vf032b with page size 256 Bytes, erase size 4 KiB, total 4 MiB *** Warning - bad CRC, using default environment

  In:    serial@10010000
  Out:   serial@10010000
  Err:   serial@10010000
  Net:   failed to get gemgxl_reset reset

Warning: ethernet@10090000 MAC addresses don't match:
Address in DT is 52:54:00:12:34:56
Address in environment is 70:b3:d5:92:f0:01
eth0: ethernet@10090000
Hit any key to stop autoboot: 0
=> sf probe
SF: Detected sst25vf032b with page size 256 Bytes, erase size 4 KiB, total 4 MiB
=> sf test 1ff000 1000
SPI flash test:
0 erase: 0 ticks, 4096000 KiB/s 32768.000 Mbps
1 check: 10 ticks, 400 KiB/s 3.200 Mbps
2 write: 170 ticks, 23 KiB/s 0.184 Mbps
3 read: 9 ticks, 444 KiB/s 3.552 Mbps
Test passed
0 erase: 0 ticks, 4096000 KiB/s 32768.000 Mbps
1 check: 10 ticks, 400 KiB/s 3.200 Mbps
2 write: 170 ticks, 23 KiB/s 0.184 Mbps
3 read: 9 ticks, 444 KiB/s 3.552 Mbps

QEMU flash: "mx66u51235f"

U-Boot 2020.10 (Jan 14 2021 - 21:55:59 +0800)

CPU: rv64imafdcsu
Model: SiFive HiFive Unleashed A00
DRAM: 2 GiB
MMC:
Loading Environment from SPIFlash... SF: Detected mx66u51235f with page size 256 Bytes, erase size 4 KiB, total 64 MiB *** Warning - bad CRC, using default environment

  In:    serial@10010000
  Out:   serial@10010000
  Err:   serial@10010000
  Net:   failed to get gemgxl_reset reset

Warning: ethernet@10090000 MAC addresses don't match:
Address in DT is 52:54:00:12:34:56
Address in environment is 70:b3:d5:92:f0:01
eth0: ethernet@10090000
Hit any key to stop autoboot: 0
=> sf probe
SF: Detected mx66u51235f with page size 256 Bytes, erase size 4 KiB, total 64 MiB => sf test 0 8000
SPI flash test:
0 erase: 1 ticks, 32000 KiB/s 256.000 Mbps
1 check: 80 ticks, 400 KiB/s 3.200 Mbps
2 write: 83 ticks, 385 KiB/s 3.080 Mbps
3 read: 79 ticks, 405 KiB/s 3.240 Mbps
Test passed
0 erase: 1 ticks, 32000 KiB/s 256.000 Mbps
1 check: 80 ticks, 400 KiB/s 3.200 Mbps
2 write: 83 ticks, 385 KiB/s 3.080 Mbps
3 read: 79 ticks, 405 KiB/s 3.240 Mbps

I am sure there will be bugs, and I have not tested all flashes affected. But I want to send out this series for an early discussion and comments. I will continue my testing.

Bin Meng (9):
hw/block: m25p80: Fix the number of dummy bytes needed for Windbond
flashes
hw/block: m25p80: Fix the number of dummy bytes needed for
Numonyx/Micron flashes
hw/block: m25p80: Fix the number of dummy bytes needed for Macronix
flashes
hw/block: m25p80: Fix the number of dummy bytes needed for Spansion
flashes
hw/block: m25p80: Support fast read for SST flashes hw/ssi: xilinx_spips: Fix generic fifo dummy cycle handling Revert "aspeed/smc: Fix number of dummy cycles for FAST_READ_4
command"
Revert "aspeed/smc: snoop SPI transfers to fake dummy cycles" hw/ssi: npcm7xx_fiu: Correct the dummy cycle emulation logic

 include/hw/ssi/aspeed_smc.h |   3 -
 hw/block/m25p80.c           | 153 ++++++++++++++++++++++++++++--------
 hw/ssi/aspeed_smc.c         | 116 +--------------------------
 hw/ssi/npcm7xx_fiu.c        |   8 +-
 hw/ssi/xilinx_spips.c       |  29 ++++++-
 5 files changed, 153 insertions(+), 156 deletions(-)
Signature · 1 line
2.25.1
author
#1 lore raw

PATCH 1/9 hw/block: m25p80: Fix the number of dummy bytes needed for Windbond flashes

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.
- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

Let's fix the mess from the flash side first. We start from a easy one, the Winbond flashes.

Per the Windbond W25Q256JV datasheet [1] instrunction set table (chapter 8.1.2, 8.1.3, 8.1.4, 8.1.5), fix the wrong number of dummy bytes needed for fast read commands.

[1] https://www.winbond.com/resource-files/w25q256jv%20spi%20revb%2009202016.pdf

Fixes fe8477052831 ("m25p80: Fix QIOR/DIOR handling for Winbond")
Fixes 3830c7a460b8 ("m25p80: Fix WINBOND fast read command handling")
Fixes cf6f1efe0b57 ("m25p80: Fast read commands family changes")

---

 hw/block/m25p80.c | 24 ++++++++++++++++++------
 1 file changed, 18 insertions(+), 6 deletions(-)
hw/block/m25p80.c +18 -6
index b744a58d1c..c947716f99 100644
@@ -875,9 +875,22 @@ static void decode_fast_read_cmd(Flash *s)
 {
     s->needed_bytes = get_addr_length(s);
     switch (get_man(s)) {
-    /* Dummy cycles - modeled with bytes writes instead of bits */
+    /* Dummy cycles - modeled with bytes writes */
     case MAN_WINBOND:
-        s->needed_bytes += 8;
+        switch (s->cmd_in_progress) {
+        case FAST_READ:
+        case FAST_READ4:
+            s->needed_bytes += 1;
+            break;
+        case DOR:
+        case DOR4:
+            s->needed_bytes += 2;
+            break;
+        case QOR:
+        case QOR4:
+            s->needed_bytes += 4;
+            break;
+        }
         break;
     case MAN_NUMONYX:
         s->needed_bytes += numonyx_extract_cfg_num_dummies(s);
@@ -906,7 +919,7 @@ static void decode_fast_read_cmd(Flash *s)
 static void decode_dio_read_cmd(Flash *s)
 {
     s->needed_bytes = get_addr_length(s);
-    /* Dummy cycles modeled with bytes writes instead of bits */
+    /* Dummy cycles modeled with bytes writes */
     switch (get_man(s)) {
     case MAN_WINBOND:
         s->needed_bytes += WINBOND_CONTINUOUS_READ_MODE_CMD_LEN;
@@ -945,11 +958,10 @@ static void decode_dio_read_cmd(Flash *s)
 static void decode_qio_read_cmd(Flash *s)
 {
     s->needed_bytes = get_addr_length(s);
-    /* Dummy cycles modeled with bytes writes instead of bits */
+    /* Dummy cycles modeled with bytes writes */
     switch (get_man(s)) {
     case MAN_WINBOND:
-        s->needed_bytes += WINBOND_CONTINUOUS_READ_MODE_CMD_LEN;
-        s->needed_bytes += 4;
+        s->needed_bytes += WINBOND_CONTINUOUS_READ_MODE_CMD_LEN + 2;
         break;
     case MAN_SPANSION:
         s->needed_bytes += SPANSION_CONTINUOUS_READ_MODE_CMD_LEN;
Signature · 1 line
2.25.1
author
#2 lore raw

PATCH 2/9 hw/block: m25p80: Fix the number of dummy bytes needed for Numonyx/Micron flashes

In reply to Bin Meng

From: Bin Meng <[email protected]>

Unfortunately the dummy cycle/bytes calculation for Numonyx/Micron flashes is still wrong, even though there were fixes before that tried to fix it.

First of all, the default number of dummy cycles is only related to the SPI protocol mode. For QSPI it is 10, otherwise it is 8.

Secondly, per the datasheet [1], it's clear that in Quad I/O or Dual I/O mode, the dummy bits show up on 4 or 2 lines.

The tricky part is the standard mode (extended mode). For such mode, the dummy bits are not like other flashes that they show up on the same lines as the address bits, but on the same lines as the data bits, so for a Quad Output Fast Read command (6Bh), the dummy bits must be sent on all the 4 IO lines. IOW, the total number of dummy bits depend on the command.

The datasheet does not state crystal clearly how many lines are used for 6Bh in the standard mode. We may only tell from figure 19 that is showing the command sequence and interpret that dummy cycles need to be on 4 lines for 6Bh.

Note as of today, both spi-nor drivers in U-Boot v2021.01 and Linux v5.10 has the wrong assumption for all flashes that dummy cycle bus width is the same as the address bits bus width, which is not true for the Numonyx/Micron flash in the standard mode.

Last if the total number of dummy bits is not multiple of 8, log an unimplemented message to notify user, and round it up. Right now the QEMU ssi_transfer() API transfers one byte each time to the flash. Leaving such as unimplemented will not cause any issue because as of today both spi-nor drivers in U-Boot and Linux have the assumption that the total number of dummy bits must be multiple of 8.

[1] https://media-www.micron.com/-/media/client/global/documents/products/

    data-sheet/nor-flash/serial-nor/n25q/n25q_512mb_1ce_3v_65nm.pdf
Fixes 23af26856606 ("hw/block/m25p80: Fix Numonyx fast read dummy cycle count")

---

 hw/block/m25p80.c | 62 +++++++++++++++++++++++++++++++++++++++++++----
 1 file changed, 57 insertions(+), 5 deletions(-)
hw/block/m25p80.c +57 -5
index c947716f99..c8cd12a6d3 100644
@@ -856,19 +856,71 @@ static uint8_t numonyx_extract_cfg_num_dummies(Flash *s)
     mode = numonyx_mode(s);
     num_dummies = extract32(s->volatile_cfg, 4, 4);
 
+    /*
+     * The default nubmer of dummy cycles is only related to the SPI
+     * protocol mode. For QSPI it is 10, otherwise it is 8.
+     */
     if (num_dummies == 0x0 || num_dummies == 0xf) {
+        num_dummies = (mode == MODE_QIO) ? 10 : 8;
+    }
+
+    /*
+     * Convert the number of dummy cycles to bytes
+     *
+     * Per the datasheet, it's clear that in Quad I/O or Dual I/O mode,
+     * the dummy bits show up on 4 or 2 lines.
+     *
+     * The tricky part is the standard mode (extended mode). For such
+     * mode, the dummy bits are not like other flashes that they show up
+     * on the same lines as the address bits, but on the same lines as
+     * the data bits, so for a Quad Output Fast Read command (6Bh), the
+     * dummy bits must be sent on all the 4 IO lines. IOW, the total
+     * number of dummy bits depend on the command.
+     *
+     * The datasheet does not state crystal clearly how many lines are
+     * used for 6Bh in the standard mode. We may only tell from figure 19
+     * that is showing the command sequence and interpret that dummy cycles
+     * need to be on 4 lines for 6Bh.
+     *
+     * Note as of today, both spi-nor drivers in U-Boot v2021.01 and Linux
+     * v5.10 has the wrong assumption for all flashes that dummy cycle bus
+     * width is the same as the address bits bus width, which is not true
+     * for the Numonyx/Micron flash in the standard mode.
+     */
+
+    if (mode == MODE_QIO) {
+        num_dummies *= 4;
+    } else if (mode == MODE_DIO) {
+        num_dummies *= 2;
+    } else {
         switch (s->cmd_in_progress) {
+        case QOR:
+        case QOR4:
         case QIOR:
         case QIOR4:
-            num_dummies = 10;
+            num_dummies *= 4;
             break;
-        default:
-            num_dummies = (mode == MODE_QIO) ? 10 : 8;
+        case DOR:
+        case DOR4:
+        case DIOR:
+        case DIOR4:
+            num_dummies *= 2;
             break;
-        }
+         }
+    }
+
+    /*
+     * If the total number of dummy bits is not multiple of 8, log an
+     * unimplemented message to notify user, and round it up.
+     */
+    if (num_dummies % 8) {
+        qemu_log_mask(LOG_UNIMP,
+                      "M25P80: the number of dummy bits is not multiple of 8");
+        num_dummies = ROUND_UP(num_dummies, 8);
     }
 
-    return num_dummies;
+    /* return the number of dummy bytes */
+    return num_dummies / 8;
 }
 
 static void decode_fast_read_cmd(Flash *s)
Signature · 1 line
2.25.1
author
#3 lore raw

PATCH 3/9 hw/block: m25p80: Fix the number of dummy bytes needed for Macronix flashes

In reply to Bin Meng

From: Bin Meng <[email protected]>

Per datasheet [1], the number of dummy cycles for Macronix flashes is configurable via two volatible bits (DC1, DC2) in a configuration register.

Do the same dummy cycle to dummy byte conversion fix as others.

[1] https://www.macronix.com/Lists/Datasheet/Attachments/7674/MX66U51235F,%201.8V,%20512Mb,%20v1.1.pdf

Fixes cf6f1efe0b57 ("m25p80: Fast read commands family changes")

---

 hw/block/m25p80.c | 21 +++++++++++----------
 1 file changed, 11 insertions(+), 10 deletions(-)
hw/block/m25p80.c +11 -10
index c8cd12a6d3..44508b3da9 100644
@@ -949,10 +949,10 @@ static void decode_fast_read_cmd(Flash *s)
         break;
     case MAN_MACRONIX:
         if (extract32(s->volatile_cfg, 6, 2) == 1) {
-            s->needed_bytes += 6;
-        } else {
-            s->needed_bytes += 8;
+            qemu_log_mask(LOG_UNIMP,
+                          "M25P80: the number of dummy bits is not multiple of 8");
         }
+        s->needed_bytes += 1;
         break;
     case MAN_SPANSION:
         s->needed_bytes += extract32(s->spansion_cr2v,
@@ -989,13 +989,14 @@ static void decode_dio_read_cmd(Flash *s)
     case MAN_MACRONIX:
         switch (extract32(s->volatile_cfg, 6, 2)) {
         case 1:
-            s->needed_bytes += 6;
-            break;
+            qemu_log_mask(LOG_UNIMP,
+                          "M25P80: the number of dummy bits is not multiple of 8");
+        /* fall-through */
         case 2:
-            s->needed_bytes += 8;
+            s->needed_bytes += 2;
             break;
         default:
-            s->needed_bytes += 4;
+            s->needed_bytes += 1;
             break;
         }
         break;
@@ -1028,13 +1029,13 @@ static void decode_qio_read_cmd(Flash *s)
     case MAN_MACRONIX:
         switch (extract32(s->volatile_cfg, 6, 2)) {
         case 1:
-            s->needed_bytes += 4;
+            s->needed_bytes += 2;
             break;
         case 2:
-            s->needed_bytes += 8;
+            s->needed_bytes += 4;
             break;
         default:
-            s->needed_bytes += 6;
+            s->needed_bytes += 3;
             break;
         }
         break;
Signature · 1 line
2.25.1
author
#4 lore raw

PATCH 4/9 hw/block: m25p80: Fix the number of dummy bytes needed for Spansion flashes

In reply to Bin Meng

From: Bin Meng <[email protected]>

Per datasheet [1], the number of dummy cycles for Spansion flashes is configurable via 4 volatible bits in a configuration register.

Do the same dummy cycle to dummy byte conversion fix as others.

[1] https://www.cypress.com/file/316171/download

---

 hw/block/m25p80.c | 43 +++++++++++++++++++++++++++++++------------
 1 file changed, 31 insertions(+), 12 deletions(-)
hw/block/m25p80.c +31 -12
index 44508b3da9..e1e5d5a76f 100644
@@ -955,10 +955,25 @@ static void decode_fast_read_cmd(Flash *s)
         s->needed_bytes += 1;
         break;
     case MAN_SPANSION:
-        s->needed_bytes += extract32(s->spansion_cr2v,
-                                    SPANSION_DUMMY_CLK_POS,
-                                    SPANSION_DUMMY_CLK_LEN
-                                    );
+        if (extract32(s->spansion_cr2v, SPANSION_DUMMY_CLK_POS,
+                      SPANSION_DUMMY_CLK_LEN) != 8) {
+            qemu_log_mask(LOG_UNIMP,
+                          "M25P80: the number of dummy bits is not multiple of 8");
+        }
+        switch (s->cmd_in_progress) {
+        case FAST_READ:
+        case FAST_READ4:
+            s->needed_bytes += 1;
+            break;
+        case DOR:
+        case DOR4:
+            s->needed_bytes += 2;
+            break;
+        case QOR:
+        case QOR4:
+            s->needed_bytes += 4;
+            break;
+        }
         break;
     default:
         break;
@@ -978,10 +993,12 @@ static void decode_dio_read_cmd(Flash *s)
         break;
     case MAN_SPANSION:
         s->needed_bytes += SPANSION_CONTINUOUS_READ_MODE_CMD_LEN;
-        s->needed_bytes += extract32(s->spansion_cr2v,
-                                    SPANSION_DUMMY_CLK_POS,
-                                    SPANSION_DUMMY_CLK_LEN
-                                    );
+        if (extract32(s->spansion_cr2v, SPANSION_DUMMY_CLK_POS,
+                      SPANSION_DUMMY_CLK_LEN) != 8) {
+            qemu_log_mask(LOG_UNIMP,
+                          "M25P80: the number of dummy bits is not multiple of 8");
+        }
+        s->needed_bytes += 2;
         break;
     case MAN_NUMONYX:
         s->needed_bytes += numonyx_extract_cfg_num_dummies(s);
@@ -1018,10 +1035,12 @@ static void decode_qio_read_cmd(Flash *s)
         break;
     case MAN_SPANSION:
         s->needed_bytes += SPANSION_CONTINUOUS_READ_MODE_CMD_LEN;
-        s->needed_bytes += extract32(s->spansion_cr2v,
-                                    SPANSION_DUMMY_CLK_POS,
-                                    SPANSION_DUMMY_CLK_LEN
-                                    );
+        if (extract32(s->spansion_cr2v, SPANSION_DUMMY_CLK_POS,
+                      SPANSION_DUMMY_CLK_LEN) != 8) {
+            qemu_log_mask(LOG_UNIMP,
+                          "M25P80: the number of dummy bits is not multiple of 8");
+        }
+        s->needed_bytes += 4;
         break;
     case MAN_NUMONYX:
         s->needed_bytes += numonyx_extract_cfg_num_dummies(s);
Signature · 1 line
2.25.1
author
#5 lore raw

PATCH 5/9 hw/block: m25p80: Support fast read for SST flashes

In reply to Bin Meng

From: Bin Meng <[email protected]>

Per SST25VF016B datasheet [1], SST flash requires a dummy byte after the address bytes. Note only SPI mode is supported by SST flashes.

[1] http://ww1.microchip.com/downloads/en/devicedoc/s71271_04.pdf

---

 hw/block/m25p80.c | 3 +++
 1 file changed, 3 insertions(+)
hw/block/m25p80.c +3 -0
index e1e5d5a76f..512af61ba5 100644
@@ -928,6 +928,9 @@ static void decode_fast_read_cmd(Flash *s)
     s->needed_bytes = get_addr_length(s);
     switch (get_man(s)) {
     /* Dummy cycles - modeled with bytes writes */
+    case MAN_SST:
+        s->needed_bytes += 1;
+        break;
     case MAN_WINBOND:
         switch (s->cmd_in_progress) {
         case FAST_READ:
Signature · 1 line
2.25.1
author
#6 lore raw

PATCH 6/9 hw/ssi: xilinx_spips: Fix generic fifo dummy cycle handling

In reply to Bin Meng

From: Bin Meng <[email protected]>

The description of the genenic command fifo register says:

When [receive, transmit, data_xfer] = [0,0,1], the [immediate_data] field represents the number of dummy cycle sent on the SPI interface.

However we should not simply use the programmed value to determine how many times ssi_transfer() needs to be called to send the dummy bytes. ssi_transfer() is used to transfer a byte on the line, not a sigle bit. Previously the m25p80 flash model wronly implemented the dummy cycles for fast read command on some flashes. Now this mess is corrected and SPI flash controllers need to be updated to do the right thing.

According to the example in the ZynqMP manual (ug1085, v2.2 [1]) we need to convert the number of dummy cycles to bytes according to the SPI mode being used, and transfer the bytes via ssi_transfer().

[1] https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf
table 24‐22, an example of Generic FIFO Contents for Quad I/O Read Command (EBh)

Fixes c95997a39de6 ("xilinx_spips: Add support for the ZynqMP Generic QSPI")

---

 hw/ssi/xilinx_spips.c | 29 ++++++++++++++++++++++++++++-
 1 file changed, 28 insertions(+), 1 deletion(-)
hw/ssi/xilinx_spips.c +28 -1
index a897034601..787de60f24 100644
@@ -191,6 +191,10 @@
     FIELD(GQSPI_GF_SNAPSHOT, EXPONENT, 9, 1)
     FIELD(GQSPI_GF_SNAPSHOT, DATA_XFER, 8, 1)
     FIELD(GQSPI_GF_SNAPSHOT, IMMEDIATE_DATA, 0, 8)
+#define GQSPI_GF_MODE_SPI     1
+#define GQSPI_GF_MODE_DSPI    2
+#define GQSPI_GF_MODE_QSPI    3
+
 #define R_GQSPI_MOD_ID        (0x1fc / 4)
 #define R_GQSPI_MOD_ID_RESET  (0x10a0000)
 
@@ -492,7 +496,30 @@ static void xlnx_zynqmp_qspips_flush_fifo_g(XlnxZynqMPQSPIPS *s)
                 }
                 s->regs[R_GQSPI_DATA_STS] = 1ul << imm;
             } else {
-                s->regs[R_GQSPI_DATA_STS] = imm;
+                /*
+                 * When [receive, transmit, data_xfer] = [0,0,1], it represents
+                 * the number of dummy cycle sent on the SPI interface. We need
+                 * to convert the number of dummy cycles to bytes according to
+                 * the SPI mode being used.
+                 *
+                 * Ref: ug1085 v2.2 (December 2020) table 24‐22, an example of
+                 *      Generic FIFO Contents for Quad I/O Read Command (EBh)
+                 */
+                if (!ARRAY_FIELD_EX32(s->regs, GQSPI_GF_SNAPSHOT, TRANSMIT) &&
+                    !ARRAY_FIELD_EX32(s->regs, GQSPI_GF_SNAPSHOT, RECIEVE)) {
+                    uint8_t spi_mode = ARRAY_FIELD_EX32(s->regs, GQSPI_GF_SNAPSHOT, SPI_MODE);
+                    if (spi_mode == GQSPI_GF_MODE_QSPI) {
+                        s->regs[R_GQSPI_DATA_STS] = ROUND_UP(imm * 4, 8) / 8;
+                    } else if (spi_mode == GQSPI_GF_MODE_DSPI) {
+                        s->regs[R_GQSPI_DATA_STS] = ROUND_UP(imm * 2, 8) / 8;
+                    } else if (spi_mode == GQSPI_GF_MODE_SPI) {
+                        s->regs[R_GQSPI_DATA_STS] = ROUND_UP(imm * 1, 8) / 8;
+                    } else {
+                        qemu_log_mask(LOG_GUEST_ERROR, "Unknown SPI MODE: 0x%x ", spi_mode);
+                    }
+                } else {
+                    s->regs[R_GQSPI_DATA_STS] = imm;
+                }
             }
         }
         /* Zero length transfer check */
Signature · 1 line
2.25.1
author
#7 lore raw

PATCH 7/9 Revert "aspeed/smc: Fix number of dummy cycles for FAST_READ_4 command"

In reply to Bin Meng

From: Bin Meng <[email protected]>

This reverts commit 7faf6f1790dddf9f3acf6ddd95f7bbc1b4a755d0.

The incorrect implementation of dummy cycles in m25p80 model is now corrected. Revert this commit.

---

 hw/ssi/aspeed_smc.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
hw/ssi/aspeed_smc.c +1 -1
index 16addee4dc..1e78b5232f 100644
@@ -802,11 +802,11 @@ static int aspeed_smc_num_dummies(uint8_t command)
     case FAST_READ:
     case DOR:
     case QOR:
-    case FAST_READ_4:
     case DOR_4:
     case QOR_4:
         return 1;
     case DIOR:
+    case FAST_READ_4:
     case DIOR_4:
         return 2;
     case QIOR:
Signature · 1 line
2.25.1
author
#8 lore raw

PATCH 8/9 Revert "aspeed/smc: snoop SPI transfers to fake dummy cycles"

In reply to Bin Meng

From: Bin Meng <[email protected]>

This reverts commit f95c4bffdc4c53b29f89762cab4adc5a43f95daf.

The incorrect implementation of dummy cycles in m25p80 model is now corrected. Revert this commit.

---

 include/hw/ssi/aspeed_smc.h |   3 -
 hw/ssi/aspeed_smc.c         | 116 +-----------------------------------
 2 files changed, 2 insertions(+), 117 deletions(-)
include/hw/ssi/aspeed_smc.h +0 -3
index 16c03fe64f..46f3abf2e7 100644
@@ -111,9 +111,6 @@ struct AspeedSMCState {
     AddressSpace dram_as;
 
     AspeedSMCFlash *flashes;
-
-    uint8_t snoop_index;
-    uint8_t snoop_dummies;
 };
 
 #endif /* ASPEED_SMC_H */
hw/ssi/aspeed_smc.c +2 -114

154 lines not on this page · raw message

Signature · 1 line
2.25.1
author
#9 lore raw

PATCH 9/9 hw/ssi: npcm7xx_fiu: Correct the dummy cycle emulation logic

In reply to Bin Meng

From: Bin Meng <[email protected]>

I believe send_dummy_bits() should also be fixed, but I really don't know how based on my pure read/guess of the codes since there is no public datasheet available for this NPCM7xx SoC.

---

 hw/ssi/npcm7xx_fiu.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
hw/ssi/npcm7xx_fiu.c +3 -5
index 5040132b07..e76fb5ad9f 100644
@@ -150,7 +150,7 @@ static uint64_t npcm7xx_fiu_flash_read(void *opaque, hwaddr addr,
     NPCM7xxFIUState *fiu = f->fiu;
     uint64_t value = 0;
     uint32_t drd_cfg;
-    int dummy_cycles;
+    int dummy_bytes;
     int i;
 
     if (fiu->active_cs != -1) {
@@ -180,10 +180,8 @@ static uint64_t npcm7xx_fiu_flash_read(void *opaque, hwaddr addr,
         break;
     }
 
-    /* Flash chip model expects one transfer per dummy bit, not byte */
-    dummy_cycles =
-        (FIU_DRD_CFG_DBW(drd_cfg) * 8) >> FIU_DRD_CFG_ACCTYPE(drd_cfg);
-    for (i = 0; i < dummy_cycles; i++) {
+    dummy_bytes = FIU_DRD_CFG_DBW(drd_cfg) >> FIU_DRD_CFG_ACCTYPE(drd_cfg);
+    for (i = 0; i < dummy_bytes; i++) {
         ssi_transfer(fiu->spi, 0);
     }
 
Signature · 1 line
2.25.1
#10 lore raw

Re: [PATCH 9/9] hw/ssi: npcm7xx_fiu: Correct the dummy cycle emulation logic

Quoted 8 lines · Bin Meng

From: Bin Meng <[email protected]>

I believe send_dummy_bits() should also be fixed, but I really don't know how based on my pure read/guess of the codes since there is no public datasheet available for this NPCM7xx SoC.

Just a quick comment before I look at the rest of the patch series:
The emulated dummy bits behavior has a lot more to do with what the m25p80 emulator seemed to expect than the actual NPCM7xx behavior. If the m25p behavior now interprets the dummy cycles the same way as the rest of the cycles, this change seems correct, but you're right that send_dummy_bits probably needs some attention as well.

I _think_ it's just a matter of turning this:

        for (j = 0; j < 8; j += bits_per_clock) {
            ssi_transfer(spi, extract32(uma_cmd, field + j, bits_per_clock));
        }

into this:

        ssi_transfer(spi, extract32(uma_cmd, field, BITS_PER_BYTE));

which might have the very nice side effect of speeding up SPI flash access quite a bit.

Thanks a lot for looking into this.

Quoted 35 lines

---

 hw/ssi/npcm7xx_fiu.c | 8 +++-----
 1 file changed, 3 insertions(+), 5 deletions(-)
hw/ssi/npcm7xx_fiu.c +3 -6
index 5040132b07..e76fb5ad9f 100644
@@ -150,7 +150,7 @@ static uint64_t npcm7xx_fiu_flash_read(void *opaque, hwaddr addr,
     NPCM7xxFIUState *fiu = f->fiu;
     uint64_t value = 0;
     uint32_t drd_cfg;
-    int dummy_cycles;
+    int dummy_bytes;
     int i;
 
     if (fiu->active_cs != -1) {
@@ -180,10 +180,8 @@ static uint64_t npcm7xx_fiu_flash_read(void *opaque, hwaddr addr,
         break;
     }
 
-    /* Flash chip model expects one transfer per dummy bit, not byte */
-    dummy_cycles =
-        (FIU_DRD_CFG_DBW(drd_cfg) * 8) >> FIU_DRD_CFG_ACCTYPE(drd_cfg);
-    for (i = 0; i < dummy_cycles; i++) {
+    dummy_bytes = FIU_DRD_CFG_DBW(drd_cfg) >> FIU_DRD_CFG_ACCTYPE(drd_cfg);
+    for (i = 0; i < dummy_bytes; i++) {
         ssi_transfer(fiu->spi, 0);
     }
 
--

2.25.1

#11 lore raw
In reply to Bin Meng
Quoted 35 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.
- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

This series fixes the mess in the m25p80 from the flash side first, followed by fixes to 3 known SPI controller models that fall into the 2nd case above.

Quoted 5 lines

Please note, I have no way to verify patch 7/8/9 because:

* There is no public datasheet available for the SoC / SPI controller
* There is no QEMU docs, or details that tell people how to boot either U-Boot or Linux kernel to verify the functionality

The Linux drivers are available in mainline but these branches are more up to date since not everything is merged :

https://github.com/openbmc/linux

u-boot :

https://github.com/openbmc/u-boot/tree/v2016.07-aspeed-openbmc (ast2400/ast2500)
https://github.com/openbmc/u-boot/tree/v2019.04-aspeed-openbmc (ast2600)

A quick intro :

https://www.qemu.org/docs/master/system/arm/aspeed.html

Quoted 5 lines

These 3 patches are very likely to be wrong. Hence I would like to ask help from the original author who wrote these SPI controller models to help testing, or completely rewrite these 3 patches to fix things. Thanks!

A quick test shows that all Aspeed machines are broken with this patchset.

Please try these command lines :

wget https://openpower.xyz/job/openbmc-build/lastSuccessfulBuild/distro=ubuntu,label=builder,target=palmetto/artifact/deploy/images/palmetto/flash-palmetto
wget https://openpower.xyz/job/openbmc-build/lastSuccessfulBuild/distro=ubuntu,label=builder,target=romulus/artifact/deploy/images/romulus/flash-romulus
wget https://openpower.xyz/job/openbmc-build/lastSuccessfulBuild/distro=ubuntu,label=builder,target=witherspoon/artifact/deploy/images/witherspoon/obmc-phosphor-image-witherspoon.ubi.mtd

qemu-system-arm -M witherspoon-bmc -nic user -drive file=obmc-phosphor-image-witherspoon.ubi.mtd,format=raw,if=mtd -nographic qemu-system-arm -M romulus-bmc -nic user -drive file=flash-romulus,format=raw,if=mtd -nographic qemu-system-arm -M palmetto-bmc -nic user -drive file=flash-palmetto,format=raw,if=mtd -nographic

The Aspeed SMC model has traces to help you in the task.

Thanks,

C.

Quoted 114 lines

Patch 6 is unvalidated with QEMU, mainly because there is no doc to tell people how to boot anything to test. But I have some confidence based on my read of the ZynqMP manual, as well as some experimental testing on a real ZCU102 board.

Other flash patches can be tested with the SiFive SPI series:
http://patchwork.ozlabs.org/project/qemu-devel/list/?series=222391

Cherry-pick patch 16 and 17 from the series above, and switch to different flash model to test with the following command:

$ qemu-system-riscv64 -nographic -M sifive_u -m 2G -smp 5 -kernel u-boot

I've picked up two for testing:

QEMU flash: "sst25vf032b"

U-Boot 2020.10 (Jan 14 2021 - 21:55:59 +0800)

CPU: rv64imafdcsu
Model: SiFive HiFive Unleashed A00
DRAM: 2 GiB
MMC:
Loading Environment from SPIFlash... SF: Detected sst25vf032b with page size 256 Bytes, erase size 4 KiB, total 4 MiB *** Warning - bad CRC, using default environment

  In:    serial@10010000
  Out:   serial@10010000
  Err:   serial@10010000
  Net:   failed to get gemgxl_reset reset

Warning: ethernet@10090000 MAC addresses don't match:
Address in DT is 52:54:00:12:34:56
Address in environment is 70:b3:d5:92:f0:01
eth0: ethernet@10090000
Hit any key to stop autoboot: 0
=> sf probe
SF: Detected sst25vf032b with page size 256 Bytes, erase size 4 KiB, total 4 MiB
=> sf test 1ff000 1000
SPI flash test:
0 erase: 0 ticks, 4096000 KiB/s 32768.000 Mbps
1 check: 10 ticks, 400 KiB/s 3.200 Mbps
2 write: 170 ticks, 23 KiB/s 0.184 Mbps
3 read: 9 ticks, 444 KiB/s 3.552 Mbps
Test passed
0 erase: 0 ticks, 4096000 KiB/s 32768.000 Mbps
1 check: 10 ticks, 400 KiB/s 3.200 Mbps
2 write: 170 ticks, 23 KiB/s 0.184 Mbps
3 read: 9 ticks, 444 KiB/s 3.552 Mbps

QEMU flash: "mx66u51235f"

U-Boot 2020.10 (Jan 14 2021 - 21:55:59 +0800)

CPU: rv64imafdcsu
Model: SiFive HiFive Unleashed A00
DRAM: 2 GiB
MMC:
Loading Environment from SPIFlash... SF: Detected mx66u51235f with page size 256 Bytes, erase size 4 KiB, total 64 MiB *** Warning - bad CRC, using default environment

  In:    serial@10010000
  Out:   serial@10010000
  Err:   serial@10010000
  Net:   failed to get gemgxl_reset reset

Warning: ethernet@10090000 MAC addresses don't match:
Address in DT is 52:54:00:12:34:56
Address in environment is 70:b3:d5:92:f0:01
eth0: ethernet@10090000
Hit any key to stop autoboot: 0
=> sf probe
SF: Detected mx66u51235f with page size 256 Bytes, erase size 4 KiB, total 64 MiB => sf test 0 8000
SPI flash test:
0 erase: 1 ticks, 32000 KiB/s 256.000 Mbps
1 check: 80 ticks, 400 KiB/s 3.200 Mbps
2 write: 83 ticks, 385 KiB/s 3.080 Mbps
3 read: 79 ticks, 405 KiB/s 3.240 Mbps
Test passed
0 erase: 1 ticks, 32000 KiB/s 256.000 Mbps
1 check: 80 ticks, 400 KiB/s 3.200 Mbps
2 write: 83 ticks, 385 KiB/s 3.080 Mbps
3 read: 79 ticks, 405 KiB/s 3.240 Mbps

I am sure there will be bugs, and I have not tested all flashes affected. But I want to send out this series for an early discussion and comments. I will continue my testing.

Bin Meng (9):
hw/block: m25p80: Fix the number of dummy bytes needed for Windbond
flashes
hw/block: m25p80: Fix the number of dummy bytes needed for
Numonyx/Micron flashes
hw/block: m25p80: Fix the number of dummy bytes needed for Macronix
flashes
hw/block: m25p80: Fix the number of dummy bytes needed for Spansion
flashes
hw/block: m25p80: Support fast read for SST flashes hw/ssi: xilinx_spips: Fix generic fifo dummy cycle handling Revert "aspeed/smc: Fix number of dummy cycles for FAST_READ_4
command"
Revert "aspeed/smc: snoop SPI transfers to fake dummy cycles" hw/ssi: npcm7xx_fiu: Correct the dummy cycle emulation logic

 include/hw/ssi/aspeed_smc.h |   3 -
 hw/block/m25p80.c           | 153 ++++++++++++++++++++++++++++--------
 hw/ssi/aspeed_smc.c         | 116 +--------------------------
 hw/ssi/npcm7xx_fiu.c        |   8 +-
 hw/ssi/xilinx_spips.c       |  29 ++++++-
 5 files changed, 153 insertions(+), 156 deletions(-)
#12 lore raw
In reply to Bin Meng
Patchew URL: https://patchew.org/QEMU/[email protected]/

Patchew URL: https://patchew.org/QEMU/[email protected]/

Hi,

This series seems to have some coding style problems. See output below for more information:

Type: series
Message-id: [email protected]
Subject: [PATCH 0/9] hw/block: m25p80: Fix the mess of dummy bytes needed for fast read commands

=== TEST SCRIPT BEGIN ===
#!/bin/bash
git rev-parse base > /dev/null || exit 0
git config --local diff.renamelimit 0
git config --local diff.renames True
git config --local diff.algorithm histogram
./scripts/checkpatch.pl --mailback base..
=== TEST SCRIPT END ===

Updating 3c8cf5a9c21ff8782164d1def7f44bd888713384
From https://github.com/patchew-project/qemu
- [tag update] patchew/[email protected] -> patchew/[email protected]
* [new tag] patchew/[email protected] -> patchew/[email protected]
Switched to a new branch 'test'
b87aded hw/ssi: npcm7xx_fiu: Correct the dummy cycle emulation logic 4518be2 Revert "aspeed/smc: snoop SPI transfers to fake dummy cycles" 6a4067a Revert "aspeed/smc: Fix number of dummy cycles for FAST_READ_4 command" e5ea744 hw/ssi: xilinx_spips: Fix generic fifo dummy cycle handling
3294942 hw/block: m25p80: Support fast read for SST flashes
50a7f9f hw/block: m25p80: Fix the number of dummy bytes needed for Spansion flashes cf6f8e1 hw/block: m25p80: Fix the number of dummy bytes needed for Macronix flashes 3925fcf hw/block: m25p80: Fix the number of dummy bytes needed for Numonyx/Micron flashes 5344168 hw/block: m25p80: Fix the number of dummy bytes needed for Windbond flashes

=== OUTPUT BEGIN ===
1/9 Checking commit 5344168de433 (hw/block: m25p80: Fix the number of dummy bytes needed for Windbond flashes) 2/9 Checking commit 3925fcf79dbc (hw/block: m25p80: Fix the number of dummy bytes needed for Numonyx/Micron flashes) 3/9 Checking commit cf6f8e145faa (hw/block: m25p80: Fix the number of dummy bytes needed for Macronix flashes) 4/9 Checking commit 50a7f9fb909b (hw/block: m25p80: Fix the number of dummy bytes needed for Spansion flashes) 5/9 Checking commit 3294942ca3a1 (hw/block: m25p80: Support fast read for SST flashes)
6/9 Checking commit e5ea74473d87 (hw/ssi: xilinx_spips: Fix generic fifo dummy cycle handling)
ERROR: line over 90 characters
#63: FILE: hw/ssi/xilinx_spips.c:510:
+ uint8_t spi_mode = ARRAY_FIELD_EX32(s->regs, GQSPI_GF_SNAPSHOT, SPI_MODE);

ERROR: line over 90 characters
#71: FILE: hw/ssi/xilinx_spips.c:518:
+ qemu_log_mask(LOG_GUEST_ERROR, "Unknown SPI MODE: 0x%x ", spi_mode);

total: 2 errors, 0 warnings, 41 lines checked

Patch 6/9 has style problems, please review. If any of these errors are false positives report them to the maintainer, see CHECKPATCH in MAINTAINERS.

7/9 Checking commit 6a4067a6a9fc (Revert "aspeed/smc: Fix number of dummy cycles for FAST_READ_4 command") 8/9 Checking commit 4518be22e1c9 (Revert "aspeed/smc: snoop SPI transfers to fake dummy cycles") 9/9 Checking commit b87aded6dc2a (hw/ssi: npcm7xx_fiu: Correct the dummy cycle emulation logic) === OUTPUT END ===

Test command exited with code: 1

The full log is available at
http://patchew.org/logs/[email protected]/testing.checkpatch/?type=message.
---
Email generated automatically by Patchew [https://patchew.org/].
Please send your feedback to [email protected]

#13 lore raw
In reply to Bin Meng

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 7 lines

- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

This series fixes the mess in the m25p80 from the flash side first,

Considering the problems solved by the solution in tree I find m25p80 pretty clean, at least I don't see any clearly better way for accurately modeling the dummy clock cycles. Counting bits instead of bytes would for example still force the controllers to mark which bits to count (when transmitting one dummy byte from a txfifo on four lines (Quad command) it generates 2 dummy clock cycles since it takes two cycles to transfer 8 bits).

Best regards,
Francisco Iglesias

Quoted 131 lines

followed by fixes to 3 known SPI controller models that fall into the 2nd case above.

Please note, I have no way to verify patch 7/8/9 because:

* There is no public datasheet available for the SoC / SPI controller
* There is no QEMU docs, or details that tell people how to boot either U-Boot or Linux kernel to verify the functionality

These 3 patches are very likely to be wrong. Hence I would like to ask help from the original author who wrote these SPI controller models to help testing, or completely rewrite these 3 patches to fix things. Thanks!

Patch 6 is unvalidated with QEMU, mainly because there is no doc to tell people how to boot anything to test. But I have some confidence based on my read of the ZynqMP manual, as well as some experimental testing on a real ZCU102 board.

Other flash patches can be tested with the SiFive SPI series:
http://patchwork.ozlabs.org/project/qemu-devel/list/?series=222391

Cherry-pick patch 16 and 17 from the series above, and switch to different flash model to test with the following command:

$ qemu-system-riscv64 -nographic -M sifive_u -m 2G -smp 5 -kernel u-boot

I've picked up two for testing:

QEMU flash: "sst25vf032b"

U-Boot 2020.10 (Jan 14 2021 - 21:55:59 +0800)

CPU: rv64imafdcsu
Model: SiFive HiFive Unleashed A00
DRAM: 2 GiB
MMC:
Loading Environment from SPIFlash... SF: Detected sst25vf032b with page size 256 Bytes, erase size 4 KiB, total 4 MiB *** Warning - bad CRC, using default environment

  In:    serial@10010000
  Out:   serial@10010000
  Err:   serial@10010000
  Net:   failed to get gemgxl_reset reset

Warning: ethernet@10090000 MAC addresses don't match:
Address in DT is 52:54:00:12:34:56
Address in environment is 70:b3:d5:92:f0:01
eth0: ethernet@10090000
Hit any key to stop autoboot: 0
=> sf probe
SF: Detected sst25vf032b with page size 256 Bytes, erase size 4 KiB, total 4 MiB
=> sf test 1ff000 1000
SPI flash test:
0 erase: 0 ticks, 4096000 KiB/s 32768.000 Mbps
1 check: 10 ticks, 400 KiB/s 3.200 Mbps
2 write: 170 ticks, 23 KiB/s 0.184 Mbps
3 read: 9 ticks, 444 KiB/s 3.552 Mbps
Test passed
0 erase: 0 ticks, 4096000 KiB/s 32768.000 Mbps
1 check: 10 ticks, 400 KiB/s 3.200 Mbps
2 write: 170 ticks, 23 KiB/s 0.184 Mbps
3 read: 9 ticks, 444 KiB/s 3.552 Mbps

QEMU flash: "mx66u51235f"

U-Boot 2020.10 (Jan 14 2021 - 21:55:59 +0800)

CPU: rv64imafdcsu
Model: SiFive HiFive Unleashed A00
DRAM: 2 GiB
MMC:
Loading Environment from SPIFlash... SF: Detected mx66u51235f with page size 256 Bytes, erase size 4 KiB, total 64 MiB *** Warning - bad CRC, using default environment

  In:    serial@10010000
  Out:   serial@10010000
  Err:   serial@10010000
  Net:   failed to get gemgxl_reset reset

Warning: ethernet@10090000 MAC addresses don't match:
Address in DT is 52:54:00:12:34:56
Address in environment is 70:b3:d5:92:f0:01
eth0: ethernet@10090000
Hit any key to stop autoboot: 0
=> sf probe
SF: Detected mx66u51235f with page size 256 Bytes, erase size 4 KiB, total 64 MiB => sf test 0 8000
SPI flash test:
0 erase: 1 ticks, 32000 KiB/s 256.000 Mbps
1 check: 80 ticks, 400 KiB/s 3.200 Mbps
2 write: 83 ticks, 385 KiB/s 3.080 Mbps
3 read: 79 ticks, 405 KiB/s 3.240 Mbps
Test passed
0 erase: 1 ticks, 32000 KiB/s 256.000 Mbps
1 check: 80 ticks, 400 KiB/s 3.200 Mbps
2 write: 83 ticks, 385 KiB/s 3.080 Mbps
3 read: 79 ticks, 405 KiB/s 3.240 Mbps

I am sure there will be bugs, and I have not tested all flashes affected. But I want to send out this series for an early discussion and comments. I will continue my testing.

Bin Meng (9):
hw/block: m25p80: Fix the number of dummy bytes needed for Windbond
flashes
hw/block: m25p80: Fix the number of dummy bytes needed for
Numonyx/Micron flashes
hw/block: m25p80: Fix the number of dummy bytes needed for Macronix
flashes
hw/block: m25p80: Fix the number of dummy bytes needed for Spansion
flashes
hw/block: m25p80: Support fast read for SST flashes hw/ssi: xilinx_spips: Fix generic fifo dummy cycle handling Revert "aspeed/smc: Fix number of dummy cycles for FAST_READ_4
command"
Revert "aspeed/smc: snoop SPI transfers to fake dummy cycles" hw/ssi: npcm7xx_fiu: Correct the dummy cycle emulation logic

 include/hw/ssi/aspeed_smc.h |   3 -
 hw/block/m25p80.c           | 153 ++++++++++++++++++++++++++++--------
 hw/ssi/aspeed_smc.c         | 116 +--------------------------
 hw/ssi/npcm7xx_fiu.c        |   8 +-
 hw/ssi/xilinx_spips.c       |  29 ++++++-
 5 files changed, 153 insertions(+), 156 deletions(-)
Signature · 1 line
2.25.1
author
#14 lore raw

Hi Francisco,

Quoted 22 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

Quoted 8 lines

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 9 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Quoted 4 lines

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

Quoted 7 lines
Quoted 7 lines

- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

Quoted 9 lines
Quoted 1 line

This series fixes the mess in the m25p80 from the flash side first,

Considering the problems solved by the solution in tree I find m25p80 pretty clean, at least I don't see any clearly better way for accurately modeling the dummy clock cycles. Counting bits instead of bytes would for example still force the controllers to mark which bits to count (when transmitting one dummy byte from a txfifo on four lines (Quad command) it generates 2 dummy clock cycles since it takes two cycles to transfer 8 bits).

SPI is a bit based protocol, not bytes. If you insist on bit modeling with the dummy cycles then you should also suggest we change all cycles (including command/addr/dummy/data phases) to be modeled with bits. That way we can accurately emulate everything, for example one potential problem like transferring 9 bit in the data phase.

However modeling everything with bit is super inefficient. My view is that we should avoid trying to support uncommon use cases (like not multiple of 8 for dummy bits) in QEMU.

Regards,
Bin

#15 lore raw

Hi Bin,

Quoted 95 lines · Bin Meng

Hi Francisco,

Quoted 22 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

Quoted 8 lines

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 9 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Quoted 4 lines

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

Quoted 7 lines
Quoted 7 lines

- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

Quoted 9 lines
Quoted 1 line

This series fixes the mess in the m25p80 from the flash side first,

Considering the problems solved by the solution in tree I find m25p80 pretty clean, at least I don't see any clearly better way for accurately modeling the dummy clock cycles. Counting bits instead of bytes would for example still force the controllers to mark which bits to count (when transmitting one dummy byte from a txfifo on four lines (Quad command) it generates 2 dummy clock cycles since it takes two cycles to transfer 8 bits).

Quoted 5 lines

SPI is a bit based protocol, not bytes. If you insist on bit modeling with the dummy cycles then you should also suggest we change all cycles (including command/addr/dummy/data phases) to be modeled with bits. That way we can accurately emulate everything, for example one potential problem like transferring 9 bit in the data phase.

I agree with this. There's really nothing special about dummy cycles. Making them special makes it super painful to implement SPI controller emulation because you have to anticipate when ssi_transfer changes semantics from byte-at-a-time to bit-at-a-time. I doubt all the SPI controllers in the tree gets it right all the time.

Quoted 3 lines

However modeling everything with bit is super inefficient. My view is that we should avoid trying to support uncommon use cases (like not multiple of 8 for dummy bits) in QEMU.

Perhaps ssi_transfer could take an additional bits parameter? That should make it possible to transfer any number of bits up to 32, while keeping the common case simple on both sides. And it would work for any SPI transfer, not just dummy cycles.

Havard

author
#16 lore raw

Hi Havard,

Quoted 105 lines · Havard Skinnemoen

Hi Bin,

Quoted 95 lines · Bin Meng

Hi Francisco,

Quoted 22 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

Quoted 8 lines

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 9 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Quoted 4 lines

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

Quoted 7 lines
Quoted 7 lines

- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

Quoted 9 lines
Quoted 1 line

This series fixes the mess in the m25p80 from the flash side first,

Considering the problems solved by the solution in tree I find m25p80 pretty clean, at least I don't see any clearly better way for accurately modeling the dummy clock cycles. Counting bits instead of bytes would for example still force the controllers to mark which bits to count (when transmitting one dummy byte from a txfifo on four lines (Quad command) it generates 2 dummy clock cycles since it takes two cycles to transfer 8 bits).

Quoted 5 lines

SPI is a bit based protocol, not bytes. If you insist on bit modeling with the dummy cycles then you should also suggest we change all cycles (including command/addr/dummy/data phases) to be modeled with bits. That way we can accurately emulate everything, for example one potential problem like transferring 9 bit in the data phase.

Quoted 6 lines

I agree with this. There's really nothing special about dummy cycles. Making them special makes it super painful to implement SPI controller emulation because you have to anticipate when ssi_transfer changes semantics from byte-at-a-time to bit-at-a-time. I doubt all the SPI controllers in the tree gets it right all the time.

Yep, it's not just painful for SPI controllers, and for the case 1 SPI controller it's impossible to snoop the data to distinguish when the dummy cycles begin.

Quoted 8 lines
Quoted 3 lines

However modeling everything with bit is super inefficient. My view is that we should avoid trying to support uncommon use cases (like not multiple of 8 for dummy bits) in QEMU.

Perhaps ssi_transfer could take an additional bits parameter? That should make it possible to transfer any number of bits up to 32, while keeping the common case simple on both sides. And it would work for any SPI transfer, not just dummy cycles.

This sounds like a good tradeoff from the emulator perspective. But I am not sure we should do this to solve the dummy cycle mess given all the default dummy cycle configurations so far match the multiple of 8 assumption.

Regards,
Bin

#17 lore raw
In reply to Bin Meng

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 22 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

Quoted 8 lines

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support for the commands is because no request has been made for them. Also there is one controller that has support.

Quoted 30 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 9 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Quoted 7 lines
Quoted 4 lines

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

static void decode_fast_read_cmd(Flash *s)

{
    uint8_t dummy_clk_cycles = 0;
    uint8_t extra_bytes;

    s->needed_bytes = get_addr_length(s);

    /* Obtain the number of dummy clock cycles needed */
    switch (get_man(s)) {
    case MAN_WINBOND:
        dummy_clk_cycles += 8;
        break;
    case MAN_NUMONYX:
        dummy_clk_cycles += numonyx_extract_cfg_num_dummies(s);
        break;
    case MAN_MACRONIX:
        if (extract32(s->volatile_cfg, 6, 2) == 1) {
            dummy_clk_cycles += 6;
        } else {
            dummy_clk_cycles += 8;
        }
        break;
    case MAN_SPANSION:
        dummy_clk_cycles += extract32(s->spansion_cr2v,
                                    SPANSION_DUMMY_CLK_POS,
                                    SPANSION_DUMMY_CLK_LEN
                                    );
        break;
    default:
        break;
    }

    if (s->model_dummy_bytes) {
        int lines = 1;

        /*
         * Expect dummy bytes from the controller so convert the dummy
         * clock cycles to dummy_bytes.
         */
        extra_bytes = convert_to_dummy_bytes(dummy_clk_count, lines);
    } else {
        /* Model individual dummy clock cycles as byte writes */
        extra_bytes = dummy_clk_cycles;
    }

    s->needed_bytes += extra_bytes;
    s->pos = 0;
    s->len = 0;
    s->state = STATE_COLLECTING_DATA;
}

Best regards,
Francisco Iglesias

Quoted 30 lines
Quoted 8 lines
Quoted 8 lines

- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

Quoted 9 lines
Quoted 1 line

This series fixes the mess in the m25p80 from the flash side first,

Considering the problems solved by the solution in tree I find m25p80 pretty clean, at least I don't see any clearly better way for accurately modeling the dummy clock cycles. Counting bits instead of bytes would for example still force the controllers to mark which bits to count (when transmitting one dummy byte from a txfifo on four lines (Quad command) it generates 2 dummy clock cycles since it takes two cycles to transfer 8 bits).

SPI is a bit based protocol, not bytes. If you insist on bit modeling with the dummy cycles then you should also suggest we change all cycles (including command/addr/dummy/data phases) to be modeled with bits. That way we can accurately emulate everything, for example one potential problem like transferring 9 bit in the data phase.

However modeling everything with bit is super inefficient. My view is that we should avoid trying to support uncommon use cases (like not multiple of 8 for dummy bits) in QEMU.

Regards,
Bin

author
#18 lore raw

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Quoted 2 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 2 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

[1] http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

Quoted 52 lines

static void decode_fast_read_cmd(Flash *s)

{
    uint8_t dummy_clk_cycles = 0;
    uint8_t extra_bytes;

    s->needed_bytes = get_addr_length(s);

    /* Obtain the number of dummy clock cycles needed */
    switch (get_man(s)) {
    case MAN_WINBOND:
        dummy_clk_cycles += 8;
        break;
    case MAN_NUMONYX:
        dummy_clk_cycles += numonyx_extract_cfg_num_dummies(s);
        break;
    case MAN_MACRONIX:
        if (extract32(s->volatile_cfg, 6, 2) == 1) {
            dummy_clk_cycles += 6;
        } else {
            dummy_clk_cycles += 8;
        }
        break;
    case MAN_SPANSION:
        dummy_clk_cycles += extract32(s->spansion_cr2v,
                                    SPANSION_DUMMY_CLK_POS,
                                    SPANSION_DUMMY_CLK_LEN
                                    );
        break;
    default:
        break;
    }

    if (s->model_dummy_bytes) {
        int lines = 1;

        /*
         * Expect dummy bytes from the controller so convert the dummy
         * clock cycles to dummy_bytes.
         */
        extra_bytes = convert_to_dummy_bytes(dummy_clk_count, lines);
    } else {
        /* Model individual dummy clock cycles as byte writes */
        extra_bytes = dummy_clk_cycles;
    }

    s->needed_bytes += extra_bytes;
    s->pos = 0;
    s->len = 0;
    s->state = STATE_COLLECTING_DATA;
}
Quoted 30 lines

Best regards,
Francisco Iglesias

Quoted 27 lines
Quoted 8 lines
Quoted 8 lines

- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

Quoted 9 lines
Quoted 1 line

This series fixes the mess in the m25p80 from the flash side first,

Considering the problems solved by the solution in tree I find m25p80 pretty clean, at least I don't see any clearly better way for accurately modeling the dummy clock cycles. Counting bits instead of bytes would for example still force the controllers to mark which bits to count (when transmitting one dummy byte from a txfifo on four lines (Quad command) it generates 2 dummy clock cycles since it takes two cycles to transfer 8 bits).

SPI is a bit based protocol, not bytes. If you insist on bit modeling with the dummy cycles then you should also suggest we change all cycles (including command/addr/dummy/data phases) to be modeled with bits. That way we can accurately emulate everything, for example one potential problem like transferring 9 bit in the data phase.

However modeling everything with bit is super inefficient. My view is that we should avoid trying to support uncommon use cases (like not multiple of 8 for dummy bits) in QEMU.

Regards,
Bin

#19 lore raw

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 7 lines
Quoted 3 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 63 lines
Quoted 3 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

Quoted 6 lines

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

In [2] instructions on how to boot u-boot/Linux is found. For building the various software components I followed the official doc in [3].

Best regards,
Francisco

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md
[3] https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/460653138/Xilinx+Open+Source+Linux

Quoted 88 lines

[1] http://patchwork.ozlabs.org/project/qemu-devel/patch/[email protected]/

Quoted 52 lines

static void decode_fast_read_cmd(Flash *s)

{
    uint8_t dummy_clk_cycles = 0;
    uint8_t extra_bytes;

    s->needed_bytes = get_addr_length(s);

    /* Obtain the number of dummy clock cycles needed */
    switch (get_man(s)) {
    case MAN_WINBOND:
        dummy_clk_cycles += 8;
        break;
    case MAN_NUMONYX:
        dummy_clk_cycles += numonyx_extract_cfg_num_dummies(s);
        break;
    case MAN_MACRONIX:
        if (extract32(s->volatile_cfg, 6, 2) == 1) {
            dummy_clk_cycles += 6;
        } else {
            dummy_clk_cycles += 8;
        }
        break;
    case MAN_SPANSION:
        dummy_clk_cycles += extract32(s->spansion_cr2v,
                                    SPANSION_DUMMY_CLK_POS,
                                    SPANSION_DUMMY_CLK_LEN
                                    );
        break;
    default:
        break;
    }

    if (s->model_dummy_bytes) {
        int lines = 1;

        /*
         * Expect dummy bytes from the controller so convert the dummy
         * clock cycles to dummy_bytes.
         */
        extra_bytes = convert_to_dummy_bytes(dummy_clk_count, lines);
    } else {
        /* Model individual dummy clock cycles as byte writes */
        extra_bytes = dummy_clk_cycles;
    }

    s->needed_bytes += extra_bytes;
    s->pos = 0;
    s->len = 0;
    s->state = STATE_COLLECTING_DATA;
}
Quoted 30 lines

Best regards,
Francisco Iglesias

Quoted 27 lines
Quoted 8 lines
Quoted 8 lines

- Dummy bytes not prepared by drivers. Drivers just tell the hardware the dummy cycle configuration via some registers, and hardware will automatically generate dummy cycles for us. Fixing the m25p80 model is not enough, and we will need to fix the SPI/QSPI models for such controllers.

Quoted 9 lines
Quoted 1 line

This series fixes the mess in the m25p80 from the flash side first,

Considering the problems solved by the solution in tree I find m25p80 pretty clean, at least I don't see any clearly better way for accurately modeling the dummy clock cycles. Counting bits instead of bytes would for example still force the controllers to mark which bits to count (when transmitting one dummy byte from a txfifo on four lines (Quad command) it generates 2 dummy clock cycles since it takes two cycles to transfer 8 bits).

SPI is a bit based protocol, not bytes. If you insist on bit modeling with the dummy cycles then you should also suggest we change all cycles (including command/addr/dummy/data phases) to be modeled with bits. That way we can accurately emulate everything, for example one potential problem like transferring 9 bit in the data phase.

However modeling everything with bit is super inefficient. My view is that we should avoid trying to support uncommon use cases (like not multiple of 8 for dummy bits) in QEMU.

Regards,
Bin

author
#20 lore raw

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Quoted 8 lines
Quoted 7 lines
Quoted 3 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 71 lines
Quoted 64 lines
Quoted 4 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

Quoted 6 lines

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

Quoted 2 lines

In [2] instructions on how to boot u-boot/Linux is found. For building the various software components I followed the official doc in [3].

I see the following QEMU commands are used to test booting U-Boot/Linux:

$ qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m 4G
-serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf -device loader,addr=0x40000000,file=Image -device
loader,addr=0x2000000,file=system.dtb

I am not sure where the system.dtb gets built from?

In [3], it mentions the Xilinx QEMU is used. And a different QEMU command is used as the example to launch U-Boot which is different from your command above.

See https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/18841606/QEMU+-+Zynq+UltraScale+MPSoC#QEMU-ZynqUltraScale+MPSoC-RunningaZynqUltraScale+U-bootImageOnXilinx'sARMQEMU

$ ./aarch64-softmmu/qemu-system-aarch64 -M arm-generic-fdt -serial
mon:stdio -serial /dev/null -display none \
  -device loader,addr=0xfd1a0104,data=0x8000000e,data-len=4 \ # Un-reset the A53
  -device loader,file=./pre-built/linux/images/bl31.elf,cpu-num=0 \ #
ARM Trusted Firmware
  -device loader,file=./pre-built/linux/images/u-boot.elf\ # The
u-boot exectuable
  -hw-dtb ./pre-built/linux/images/zynqmp-qemu-arm.dtb # HW Device
Tree that QEMU uses to generate the model

It is using a machine called "arm-generic-fdt", but in the mainline QEMU there is no such machine called "arm-generic-fdt".

Quoted 8 lines

Best regards,
Francisco

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md
[3] https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/460653138/Xilinx+Open+Source+Linux

Regards,
Bin

#21 lore raw

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 24 lines
Quoted 9 lines
Quoted 8 lines
Quoted 4 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Quoted 90 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 71 lines
Quoted 64 lines
Quoted 4 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

Quoted 6 lines

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

Quoted 2 lines

In [2] instructions on how to boot u-boot/Linux is found. For building the various software components I followed the official doc in [3].

Quoted 8 lines

I see the following QEMU commands are used to test booting U-Boot/Linux:

$ qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m 4G
-serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf -device loader,addr=0x40000000,file=Image -device
loader,addr=0x2000000,file=system.dtb

I am not sure where the system.dtb gets built from?

It is the instructions in [2] to look into. 'system.dtb' is the kernel dtb for zcu102 ([2] has been fixed). I created [2] purely for you, so respectfully I will ask you to try a little first before asking for further guidance.

Best regards,
Francisco Iglesias

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md

Quoted 31 lines

In [3], it mentions the Xilinx QEMU is used. And a different QEMU command is used as the example to launch U-Boot which is different from your command above.

See https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/18841606/QEMU+-+Zynq+UltraScale+MPSoC#QEMU-ZynqUltraScale+MPSoC-RunningaZynqUltraScale+U-bootImageOnXilinx'sARMQEMU

$ ./aarch64-softmmu/qemu-system-aarch64 -M arm-generic-fdt -serial
mon:stdio -serial /dev/null -display none \
  -device loader,addr=0xfd1a0104,data=0x8000000e,data-len=4 \ # Un-reset the A53
  -device loader,file=./pre-built/linux/images/bl31.elf,cpu-num=0 \ #
ARM Trusted Firmware
  -device loader,file=./pre-built/linux/images/u-boot.elf\ # The
u-boot exectuable
  -hw-dtb ./pre-built/linux/images/zynqmp-qemu-arm.dtb # HW Device
Tree that QEMU uses to generate the model

It is using a machine called "arm-generic-fdt", but in the mainline QEMU there is no such machine called "arm-generic-fdt".

Quoted 8 lines

Best regards,
Francisco

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md
[3] https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/460653138/Xilinx+Open+Source+Linux

Regards,
Bin

author
#22 lore raw

Hi Francisco,

Quoted 98 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Quoted 13 lines

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

Quoted 30 lines
Quoted 24 lines
Quoted 9 lines
Quoted 8 lines
Quoted 4 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

Quoted 81 lines
Quoted 81 lines
Quoted 71 lines
Quoted 64 lines
Quoted 4 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

Quoted 6 lines

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

Quoted 2 lines

In [2] instructions on how to boot u-boot/Linux is found. For building the various software components I followed the official doc in [3].

I see the following QEMU commands are used to test booting U-Boot/Linux:

$ qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m 4G
-serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf -device loader,addr=0x40000000,file=Image -device
loader,addr=0x2000000,file=system.dtb
Quoted 6 lines
Quoted 1 line

I am not sure where the system.dtb gets built from?

It is the instructions in [2] to look into. 'system.dtb' is the kernel dtb for zcu102 ([2] has been fixed). I created [2] purely for you, so respectfully I will ask you to try a little first before asking for further guidance.

I tried, but no success. I removed the "-device loader" part for loading kernel image and the device tree, and only focused on booting U-Boot.

The ATF bl31.elf was built from
https://github.com/ARM-software/arm-trusted-firmware, by following build instructions at
https://trustedfirmware-a.readthedocs.io/en/latest/plat/xilinx-zynqmp.html.
U-Boot was built from the upstream U-Boot.

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.4(release):v2.4-228-g337e493
NOTICE:  BL31: Built : 21:18:14, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

I also tried the Xilinx fork of ATF from
https://github.com/Xilinx/arm-trusted-firmware, by following build instructions at
https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/18842305/Build+ARM+Trusted+Firmware+ATF

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.2(release):xilinx-v2020.2
NOTICE:  BL31: Built : 21:52:38, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

Then I tried to build a U-Boot from the Xilinx fork at
https://github.com/Xilinx/u-boot-xlnx/, still no success.

Quoted 7 lines

Best regards,
Francisco Iglesias

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md

Regards,
Bin

#23 lore raw

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 98 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Quoted 13 lines

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Quoted 55 lines
Quoted 31 lines
Quoted 25 lines
Quoted 10 lines
Quoted 9 lines
Quoted 5 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Now that above has is cleared out, and since I know you are heavily loaded with other higher prio tasks, lets wait for the maintainers to also have a look into the RFC (understandibly this can take some time due to that they also are heavily loaded).

Best regards,
Francisco Iglesias

Quoted 138 lines
Quoted 81 lines
Quoted 81 lines
Quoted 71 lines
Quoted 64 lines
Quoted 4 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

Quoted 6 lines

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

Quoted 2 lines

In [2] instructions on how to boot u-boot/Linux is found. For building the various software components I followed the official doc in [3].

I see the following QEMU commands are used to test booting U-Boot/Linux:

$ qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m 4G
-serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf -device loader,addr=0x40000000,file=Image -device
loader,addr=0x2000000,file=system.dtb
Quoted 6 lines
Quoted 1 line

I am not sure where the system.dtb gets built from?

It is the instructions in [2] to look into. 'system.dtb' is the kernel dtb for zcu102 ([2] has been fixed). I created [2] purely for you, so respectfully I will ask you to try a little first before asking for further guidance.

I tried, but no success. I removed the "-device loader" part for loading kernel image and the device tree, and only focused on booting U-Boot.

The ATF bl31.elf was built from
https://github.com/ARM-software/arm-trusted-firmware, by following build instructions at
https://trustedfirmware-a.readthedocs.io/en/latest/plat/xilinx-zynqmp.html.
U-Boot was built from the upstream U-Boot.

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.4(release):v2.4-228-g337e493
NOTICE:  BL31: Built : 21:18:14, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

I also tried the Xilinx fork of ATF from
https://github.com/Xilinx/arm-trusted-firmware, by following build instructions at
https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/18842305/Build+ARM+Trusted+Firmware+ATF

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.2(release):xilinx-v2020.2
NOTICE:  BL31: Built : 21:52:38, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

Then I tried to build a U-Boot from the Xilinx fork at
https://github.com/Xilinx/u-boot-xlnx/, still no success.

Quoted 7 lines

Best regards,
Francisco Iglesias

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md

Regards,
Bin

author
#24 lore raw

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 63 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

Quoted 6 lines

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Quoted 5 lines

Now that above has is cleared out, and since I know you are heavily loaded with other higher prio tasks, lets wait for the maintainers to also have a look into the RFC (understandibly this can take some time due to that they also are heavily loaded).

Yes, maintainers are pretty much silent on this topic.

However may I ask you to provide more details on my questions below on booting U-Boot/Linux with the QEMU?

You can post patches to add documentation for zynqmp in docs/system/arm, or once I get a working instructions, I could do that too. Much appreciated.

Quoted 139 lines

Best regards,
Francisco Iglesias

Quoted 134 lines
Quoted 81 lines
Quoted 81 lines
Quoted 71 lines
Quoted 64 lines
Quoted 4 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

Quoted 6 lines

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

Quoted 2 lines

In [2] instructions on how to boot u-boot/Linux is found. For building the various software components I followed the official doc in [3].

I see the following QEMU commands are used to test booting U-Boot/Linux:

$ qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m 4G
-serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf -device loader,addr=0x40000000,file=Image -device
loader,addr=0x2000000,file=system.dtb
Quoted 6 lines
Quoted 1 line

I am not sure where the system.dtb gets built from?

It is the instructions in [2] to look into. 'system.dtb' is the kernel dtb for zcu102 ([2] has been fixed). I created [2] purely for you, so respectfully I will ask you to try a little first before asking for further guidance.

I tried, but no success. I removed the "-device loader" part for loading kernel image and the device tree, and only focused on booting U-Boot.

The ATF bl31.elf was built from
https://github.com/ARM-software/arm-trusted-firmware, by following build instructions at
https://trustedfirmware-a.readthedocs.io/en/latest/plat/xilinx-zynqmp.html.
U-Boot was built from the upstream U-Boot.

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.4(release):v2.4-228-g337e493
NOTICE:  BL31: Built : 21:18:14, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

I also tried the Xilinx fork of ATF from
https://github.com/Xilinx/arm-trusted-firmware, by following build instructions at
https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/18842305/Build+ARM+Trusted+Firmware+ATF

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.2(release):xilinx-v2020.2
NOTICE:  BL31: Built : 21:52:38, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

Then I tried to build a U-Boot from the Xilinx fork at
https://github.com/Xilinx/u-boot-xlnx/, still no success.

Quoted 6 lines

Best regards,
Francisco Iglesias

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md

Regards,
Bin

#25 lore raw

Dear Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 63 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

Quoted 6 lines

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Above transfer is already handled in the model, and since it will not change it will still work afterwards.

About below, sure I'll provide some doc once I get some time over.

Best regards,
Francisco Iglesias

Quoted 158 lines
Quoted 6 lines

Now that above has is cleared out, and since I know you are heavily loaded with other higher prio tasks, lets wait for the maintainers to also have a look into the RFC (understandibly this can take some time due to that they also are heavily loaded).

Yes, maintainers are pretty much silent on this topic.

However may I ask you to provide more details on my questions below on booting U-Boot/Linux with the QEMU?

You can post patches to add documentation for zynqmp in docs/system/arm, or once I get a working instructions, I could do that too. Much appreciated.

Quoted 139 lines

Best regards,
Francisco Iglesias

Quoted 134 lines
Quoted 81 lines
Quoted 81 lines
Quoted 71 lines
Quoted 64 lines
Quoted 4 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

Quoted 6 lines

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

Quoted 2 lines

In [2] instructions on how to boot u-boot/Linux is found. For building the various software components I followed the official doc in [3].

I see the following QEMU commands are used to test booting U-Boot/Linux:

$ qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m 4G
-serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf -device loader,addr=0x40000000,file=Image -device
loader,addr=0x2000000,file=system.dtb
Quoted 6 lines
Quoted 1 line

I am not sure where the system.dtb gets built from?

It is the instructions in [2] to look into. 'system.dtb' is the kernel dtb for zcu102 ([2] has been fixed). I created [2] purely for you, so respectfully I will ask you to try a little first before asking for further guidance.

I tried, but no success. I removed the "-device loader" part for loading kernel image and the device tree, and only focused on booting U-Boot.

The ATF bl31.elf was built from
https://github.com/ARM-software/arm-trusted-firmware, by following build instructions at
https://trustedfirmware-a.readthedocs.io/en/latest/plat/xilinx-zynqmp.html.
U-Boot was built from the upstream U-Boot.

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.4(release):v2.4-228-g337e493
NOTICE:  BL31: Built : 21:18:14, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

I also tried the Xilinx fork of ATF from
https://github.com/Xilinx/arm-trusted-firmware, by following build instructions at
https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/18842305/Build+ARM+Trusted+Firmware+ATF

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.2(release):xilinx-v2020.2
NOTICE:  BL31: Built : 21:52:38, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

Then I tried to build a U-Boot from the Xilinx fork at
https://github.com/Xilinx/u-boot-xlnx/, still no success.

Quoted 6 lines

Best regards,
Francisco Iglesias

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md

Regards,
Bin

#26 lore raw
In reply to Bin Meng

Hi Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 63 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

Quoted 6 lines

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

Best regards,
Francisco Iglesias

Quoted 158 lines
Quoted 6 lines

Now that above has is cleared out, and since I know you are heavily loaded with other higher prio tasks, lets wait for the maintainers to also have a look into the RFC (understandibly this can take some time due to that they also are heavily loaded).

Yes, maintainers are pretty much silent on this topic.

However may I ask you to provide more details on my questions below on booting U-Boot/Linux with the QEMU?

You can post patches to add documentation for zynqmp in docs/system/arm, or once I get a working instructions, I could do that too. Much appreciated.

Quoted 139 lines

Best regards,
Francisco Iglesias

Quoted 134 lines
Quoted 81 lines
Quoted 81 lines
Quoted 71 lines
Quoted 64 lines
Quoted 4 lines

for the commands is because no request has been made for them. Also there is one controller that has support.

Definitely it's not "no request". Nearly all SPI flashes support the Fast Read (0Bh) command today, and 0Bh requires a dummy cycle. This is "seriously broken" for those case 1 type controllers because they cannot read anything from the m25p80 model at all. Unless the guest software being tested only uses Read (03h) command which is not affected. But I can't find a software that uses Read instead of Fast Read.

Quoted 34 lines
Quoted 34 lines

The issue you pointed out that we require the total number of dummy bits should be multiple of 8 is true, that's why I added the unimplemented log message in this series (patch 2/3/4) to warn users if this expectation is not met. However this will not cause any issue when running U-Boot or Linux, because both spi-nor drivers expect the same assumption as we do here.

See U-Boot spi_nor_read_data() and Linux spi_nor_spimem_read_data(), there is a logic to calculate the dummy bytes needed for fast read command:

    /* convert the dummy cycles to the number of bytes */
    op.dummy.nbytes = (nor->read_dummy * op.dummy.buswidth) / 8;

Note the default dummy cycles configuration for all flashes I have looked into as of today, meets the multiple of 8 assumption. On some flashes the dummy cycle number is configurable, and if it's been configured to be an odd value, it would not work on U-Boot/Linux in the first place.

Quoted 13 lines
Quoted 8 lines

Things get complicated when interacting with different SPI or QSPI flash controllers. There are major two cases:

- Dummy bytes prepared by drivers, and wrote to the controller fifo. For such case, driver will calculate the correct number of dummy bytes and write them into the tx fifo. Fixing the m25p80 model will fix flashes working with such controllers.

Above can be fixed while still keeping the detailed dummy cycle implementation inside m25p80. Perhaps one of the following could be looked into: configurating the amount, letting the spi ctrl fetch the amount from m25p80 or by inheriting some functionality handling this in the SPI controller. Or a mixture of above.

Quoted 6 lines
Quoted 2 lines

Please send patches to explain this in detail how this is going to work. I am open to all possible solutions.

In that case I suggest that you instead try with a device property 'model_dummy_bytes' used to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80. Below is an example on how to modify the

No this is wrong in my view. This is not like a DMA vs. PIO handling.

Quoted 4 lines

decode_fast_read_cmd function (the other commands requiring dummy clock cycles can follow a similar pattern). This way the fifo mode will be able to work the way you desire while also keeping the current functionality intact. Suddenly removing functionality (features) will take users by surprise.

I don't think we are removing any features. This is a fix to make the model to be used by any SPI controllers.

Quoted 6 lines

As I pointed out, both U-Boot and Linux have the multiple of 8 assumption for the dummy bit, which is the default configuration for all flashes I have looked into so far. Can you please comment what use case you want to support? I requested a U-Boot/Linux kernel testing in the previous SST thread [1] against Xilinx GQSPI but there was no response.

Quoted 2 lines

In [2] instructions on how to boot u-boot/Linux is found. For building the various software components I followed the official doc in [3].

I see the following QEMU commands are used to test booting U-Boot/Linux:

$ qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m 4G
-serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf -device loader,addr=0x40000000,file=Image -device
loader,addr=0x2000000,file=system.dtb
Quoted 6 lines
Quoted 1 line

I am not sure where the system.dtb gets built from?

It is the instructions in [2] to look into. 'system.dtb' is the kernel dtb for zcu102 ([2] has been fixed). I created [2] purely for you, so respectfully I will ask you to try a little first before asking for further guidance.

I tried, but no success. I removed the "-device loader" part for loading kernel image and the device tree, and only focused on booting U-Boot.

The ATF bl31.elf was built from
https://github.com/ARM-software/arm-trusted-firmware, by following build instructions at
https://trustedfirmware-a.readthedocs.io/en/latest/plat/xilinx-zynqmp.html.
U-Boot was built from the upstream U-Boot.

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.4(release):v2.4-228-g337e493
NOTICE:  BL31: Built : 21:18:14, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

I also tried the Xilinx fork of ATF from
https://github.com/Xilinx/arm-trusted-firmware, by following build instructions at
https://xilinx-wiki.atlassian.net/wiki/spaces/A/pages/18842305/Build+ARM+Trusted+Firmware+ATF

$ ./qemu-system-aarch64 -M xlnx-zcu102,secure=on,virtualization=on -m
4G -serial stdio -display none -device loader,file=u-boot.elf -kernel
bl31.elf
ERROR:   Incorrect XILINX IDCODE 0x0, maskid 0x4600093
NOTICE:  ATF running on XCZUUNKN/silicon v1/RTL0.0 at 0xfffea000
NOTICE:  BL31: v2.2(release):xilinx-v2020.2
NOTICE:  BL31: Built : 21:52:38, Jan 20 2021
ERROR:   BL31: Platform Management API version error. Expected: v1.1 -
Found: v0.0
ERROR:   Error initializing runtime service sip_svc

Then I tried to build a U-Boot from the Xilinx fork at
https://github.com/Xilinx/u-boot-xlnx/, still no success.

Quoted 6 lines

Best regards,
Francisco Iglesias

[1] qemu/docs/system/deprecated.rst
[2] https://github.com/franciscoIglesias/qemu-cmdline/blob/master/xlnx-zcu102-atf-u-boot-linux.md

Regards,
Bin

author
#27 lore raw
Quoted 211 lines · Francisco Iglesias

Hi Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up bytes are expected to be received after it receives a command. For example, depending on the address mode, either 3-byte address or 4-byte address is needed.

For fast read family commands, some dummy cycles are required after sending the address bytes, and the dummy cycles need to be counted in s->needed_bytes. This is where the mess began.

Quoted 6 lines

As the variable name (needed_bytes) indicates, the unit is in byte. It is not in bit, or cycle. However for some reason the model has been using the number of dummy cycles for s->needed_bytes. The right approach is to convert the number of dummy cycles to bytes based on the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 69 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 6 lines
Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

Regards,
Bin

#28 lore raw
Quoted 328 lines · Bin Meng
Quoted 47 lines · Francisco Iglesias

Hi Bin,

Quoted 43 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Dear Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 31 lines · Francisco Iglesias

Hi Bin,

Quoted 27 lines · Bin Meng

Hi Francisco,

Quoted 23 lines · Francisco Iglesias

Hi Bin,

Quoted 19 lines · Bin Meng

Hi Francisco,

Quoted 15 lines · Francisco Iglesias

Hi Bin,

Quoted 11 lines · Bin Meng

Hi Francisco,

Quoted 7 lines · Francisco Iglesias

Hi Bin,

Quoted 3 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate

how many follow-up

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

bytes are expected to be received after it

receives a command. For

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

example, depending on the address mode, either

3-byte address or

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

4-byte address is needed.

For fast read family commands, some dummy cycles

are required after

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

sending the address bytes, and the dummy cycles

need to be counted

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the

unit is in byte.

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

It is not in bit, or cycle. However for some

reason the model has

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

been using the number of dummy cycles for

s->needed_bytes. The right

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

approach is to convert the number of dummy cycles

to bytes based on

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

the SPI protocol, for example, 6 dummy cycles for

the Fast Read Quad

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

I/O (EBh) should be converted to 3 bytes per the

formula (6 * 4 / 8).

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

While not being the original implementor I must

assume that above solution was

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

considered but not chosen by the developers due to

it is inaccuracy (it

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

wouldn't be possible to model exacly 6 dummy cycles,

only a multiple of 8,

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

meaning that if the controller is wrongly programmed

to generate 7 the error

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

wouldn't be caught and the controller will still be

considered "correct"). Now

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

that we have this detail in the implementation I'm

in favor of keeping it, this

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

also because the detail is already in use for

catching exactly above error.

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 1 line

I found no clue from the commit message that my

proposed solution here

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

was ever considered, otherwise all SPI controller

models supporting

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

software generation should have been found out

seriously broken long

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 1 line

time ago!

The controllers you are referring to might lack support

for commands requiring

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

dummy clock cycles but I really hope they work with the

other commands? If so I

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

I am not sure why you view dummy clock cycles as something

special

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

that needs some special support from the SPI controller.

For the case

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

1 controller, it's nothing special from the controller

perspective,

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

just like sending out a command, or address bytes, or

data. The

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

controller just shifts data bit by bit from its tx fifo

and that's it.

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

In the Xilinx GQSPI controller case, the dummy cycles can

either be

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

sent via a regular data (the case 1 controller) in the tx

fifo, or

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

automatically generated (case 2 controller) by the

hardware.

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

Ok, I'll try to explain my view point a little differently.

For that we also

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

need to keep in mind that QEMU models HW, and any binary

that runs on a HW

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

board supported in QEMU should ideally run on that board

inside QEMU aswell

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

(this can be a bare metal application equaly well as a

modified u-boot/Linux

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

using SPI commands with a non multiple of 8 number of dummy

clock cycles).

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

Once functionality has been introduced into QEMU it is not

easy to know which

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

intentional or untentional features provided by the

functionality are being

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

used by users. One of the (perhaps not well known) features

I'm aware of that

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

is in use and is provided by the accurate dummy clock cycle

modeling inside

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

m25p80 is the be ability to test drivers accurately

regarding the dummy clock

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

cycles (even when using commands with a non-multiple of 8

number of dummy clock

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

cycles), but there might be others aswell. So by removing

this functionality

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

above use case will brake, this since those test will not be

reliable.

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

Furthermore, since users tend to be creative it is not

possible to know if

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

there are other use cases that will be affected. This means

that in case [1]

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

needs to be followed the safe path is to add functionality

instead of removing.

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 1 line

Luckily it also easier in this case, see below.

I understand there might be users other than U-Boot/Linux that

use an

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

odd number of dummy bits (not multiple of 8). If your concern

was

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes

in the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 1 line

m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this

functionality has been

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

in tree for while, users have found known or unknown features

that got

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

introduced by it. By removing the functionality (and the

known/uknown features)

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

we are riscing to brake our user's use cases (currently I'm

aware of one

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

feature/use case but it is not unlikely that there are more).

[1] states that

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

"In general features are intended to be supported indefinitely

once introduced

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

into QEMU", to me that makes very much sense because the

opposite would mean

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

that we were not reliable. So in case [1] needs to be honored it

looks to be

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

safer to add functionality instead of removing (and riscing the

removal of use

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

cases/features). Luckily I still believe in this case that it

will be easier to

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

go forward (even if I also agree on what you are saying below

about what I

Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 7 lines
Quoted 2 lines

proposed).

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both

the case 1

Quoted 1 line
Quoted 1 line
Quoted 1 line

controller (using a txfifo) and the case 2 controller (no txfifo, as

GQSPI).

Quoted 1 line
Quoted 1 line
Quoted 1 line

Because of this, I, with all respect, will have to disagree that

this is buggy.

Quoted 10 lines
Quoted 10 lines

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

don't think it is fair to call them 'seriously broken'

(and else we should

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

probably let the maintainers know about it). Most likely

the lack of support

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

I called it "seriously broken" because current

implementation only

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

considered one type of SPI controllers while completely

ignoring the

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 1 line

other type.

If we change view and see this from the perspective of

m25p80, it models the

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

commands a certain way and provides an API that the SPI

controllers need to

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

implement for interacting with it. It is true that there are

SPI controllers

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

referred to above that do not support the portion of that

API that corresponds

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

to commands with dummy clock cycles, but I don't think it is

true that this is

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

broken since there is also one SPI controller that has a

working implementation

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

of m25p80's full API also when transfering through a tx fifo

(use case 1). But

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

as mentioned above, by doing a minor extension and

improvement to m25p80's API

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

and allow for toggling the accuracy from dummy clock cycles

to dummy bytes [1]

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

will still be honored as in the same time making it possible

to have full

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

support for the API in the SPI controllers that currently do

not (please reread

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

the proposal in my previous reply that attempts to do this).

I myself see this

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

as win/win situation, also because no controller should need

modifications.

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 1 line

I am afraid your proposal does not work. Your proposed new

device

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

property 'model_dummy_bytes' to select to convert the accurate

dummy

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

clock cycle count to dummy bytes inside m25p80, is hard to

justify as

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

a property to the flash itself, as the behavior is tightly

coupled to

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 1 line

how the SPI controller works.

I agree on above. I decided though that instead of posting

sample code in here

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

I'll post an RFC with hopefully an improved proposal. I'll cc

you. About below,

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

Xilinx ZynqMP GQSPI should not need any modication in a first

step.

Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 6 lines
Quoted 1 line

Wait, (see below)

Quoted 2 lines
Quoted 2 lines

Please take a look at the Xilinx GQSPI controller, which

supports both

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

use cases, that the dummy cycles can be transferred via tx

fifo, or

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

generated by the controller automatically. Please read the

example

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

given in:

table 24‐22, an example of Generic FIFO Contents for Quad

I/O Read

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

Command (EBh)

in

https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines
Quoted 2 lines

If you choose to set the m25p80 device property

'model_dummy_bytes' to

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

true when working with the Xilinx GQSPI controller, you are

bound to

Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line

only allow guest software to use tx fifo to transfer the dummy

cycles,

Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 4 lines
Quoted 2 lines
Quoted 2 lines

and this is wrong.

You missed this part. I looked at your RFC, and as I mentioned

above

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines

your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy

cycle

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 2 lines

generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU

transfers from 2

Quoted 1 line
Quoted 1 line
Quoted 1 line

locations in the file, in 1 location the transfer referred to above

is done, in

Quoted 1 line
Quoted 1 line
Quoted 1 line

another location the transfer through the txfifo is done. The

location where

Quoted 1 line
Quoted 1 line
Quoted 1 line

transfer referred to above is done will not need any modifications

(and will

Quoted 8 lines
Quoted 6 lines
Quoted 1 line

thus work equally well as it does currently).

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Quoted 7 lines

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

Francisco really knows this stuff better than me....
I would tend to agree that it's unfortunate to model things in cycles, if we could abstract things at a higher level that would be nice. Without breaking existing use-cases.
Francisco, is it impossible to bring up the abstraction level to bytes and keep existing use-cases?

We have a bunch of test-cases, We'll publish some of them in source code, others we can't publish since they use proprietary SW we're not allowed to publish at all, but we can run tests and Ack if things work.

Best regards,
Edgar

#29 lore raw

Hello Edgar,

Quoted 342 lines · Edgar E. Iglesias

On Mon, Feb 8, 2021 at 3:42 PM Bin Meng <[email protected]> wrote:

On Thu, Jan 21, 2021 at 10:18 PM Francisco Iglesias <[email protected]> wrote:
>
> Hi Bin,
>
> On [2021 Jan 21] Thu 16:59:51, Bin Meng wrote:
> > Hi Francisco,
> >
> > On Thu, Jan 21, 2021 at 4:50 PM Francisco Iglesias
> > <[email protected]> wrote:
> > >
> > > Dear Bin,
> > >
> > > On [2021 Jan 20] Wed 22:20:25, Bin Meng wrote:
> > > > Hi Francisco,
> > > >
> > > > On Tue, Jan 19, 2021 at 9:01 PM Francisco Iglesias
> > > > <[email protected]> wrote:
> > > > >
> > > > > Hi Bin,
> > > > >
> > > > > On [2021 Jan 18] Mon 20:32:19, Bin Meng wrote:
> > > > > > Hi Francisco,
> > > > > >
> > > > > > On Mon, Jan 18, 2021 at 6:06 PM Francisco Iglesias > > > > > > <[email protected]> wrote:
> > > > > > >
> > > > > > > Hi Bin,
> > > > > > >
> > > > > > > On [2021 Jan 15] Fri 22:38:18, Bin Meng wrote:
> > > > > > > > Hi Francisco,
> > > > > > > >
> > > > > > > > On Fri, Jan 15, 2021 at 8:26 PM Francisco Iglesias > > > > > > > > <[email protected]> wrote:
> > > > > > > > >
> > > > > > > > > Hi Bin,
> > > > > > > > >
> > > > > > > > > On [2021 Jan 15] Fri 10:07:52, Bin Meng wrote:
> > > > > > > > > > Hi Francisco,
> > > > > > > > > >
> > > > > > > > > > On Fri, Jan 15, 2021 at 2:13 AM Francisco Iglesias > > > > > > > > > > <[email protected]> wrote:
> > > > > > > > > > >
> > > > > > > > > > > Hi Bin,
> > > > > > > > > > >
> > > > > > > > > > > On [2021 Jan 14] Thu 23:08:53, Bin Meng wrote:
> > > > > > > > > > > > From: Bin Meng <[email protected]> > > > > > > > > > > > >
> > > > > > > > > > > > The m25p80 model uses s->needed_bytes to indicate how many follow-up
> > > > > > > > > > > > bytes are expected to be received after it receives a command. For
> > > > > > > > > > > > example, depending on the address mode, either 3-byte address or
> > > > > > > > > > > > 4-byte address is needed.
> > > > > > > > > > > >
> > > > > > > > > > > > For fast read family commands, some dummy cycles are required after
> > > > > > > > > > > > sending the address bytes, and the dummy cycles need to be counted
> > > > > > > > > > > > in s->needed_bytes. This is where the mess began.
> > > > > > > > > > > >
> > > > > > > > > > > > As the variable name (needed_bytes) indicates, the unit is in byte.
> > > > > > > > > > > > It is not in bit, or cycle. However for some reason the model has
> > > > > > > > > > > > been using the number of dummy cycles for s->needed_bytes. The right
> > > > > > > > > > > > approach is to convert the number of dummy cycles to bytes based on
> > > > > > > > > > > > the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad
> > > > > > > > > > > > I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).
> > > > > > > > > > >
> > > > > > > > > > > While not being the original implementor I must assume that above solution was
> > > > > > > > > > > considered but not chosen by the developers due to it is inaccuracy (it
> > > > > > > > > > > wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8,
> > > > > > > > > > > meaning that if the controller is wrongly programmed to generate 7 the error
> > > > > > > > > > > wouldn't be caught and the controller will still be considered "correct"). Now
> > > > > > > > > > > that we have this detail in the implementation I'm in favor of keeping it, this
> > > > > > > > > > > also because the detail is already in use for catching exactly above error.
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > I found no clue from the commit message that my proposed solution here
> > > > > > > > > > was ever considered, otherwise all SPI controller models supporting
> > > > > > > > > > software generation should have been found out seriously broken long
> > > > > > > > > > time ago!
> > > > > > > > >
> > > > > > > > >
> > > > > > > > > The controllers you are referring to might lack support for commands requiring
> > > > > > > > > dummy clock cycles but I really hope they work with the other commands? If so I
> > > > > > > >
> > > > > > > > I am not sure why you view dummy clock cycles as something special
> > > > > > > > that needs some special support from the SPI controller. For the case
> > > > > > > > 1 controller, it's nothing special from the controller perspective,
> > > > > > > > just like sending out a command, or address bytes, or data. The
> > > > > > > > controller just shifts data bit by bit from its tx fifo and that's it.
> > > > > > > > In the Xilinx GQSPI controller case, the dummy cycles can either be
> > > > > > > > sent via a regular data (the case 1 controller) in the tx fifo, or
> > > > > > > > automatically generated (case 2 controller) by the hardware.
> > > > > > >
> > > > > > > Ok, I'll try to explain my view point a little differently. For that we also
> > > > > > > need to keep in mind that QEMU models HW, and any binary that runs on a HW
> > > > > > > board supported in QEMU should ideally run on that board inside QEMU aswell
> > > > > > > (this can be a bare metal application equaly well as a modified u-boot/Linux
> > > > > > > using SPI commands with a non multiple of 8 number of dummy clock cycles).
> > > > > > >
> > > > > > > Once functionality has been introduced into QEMU it is not easy to know which
> > > > > > > intentional or untentional features provided by the functionality are being
> > > > > > > used by users. One of the (perhaps not well known) features I'm aware of that
> > > > > > > is in use and is provided by the accurate dummy clock cycle modeling inside
> > > > > > > m25p80 is the be ability to test drivers accurately regarding the dummy clock
> > > > > > > cycles (even when using commands with a non-multiple of 8 number of dummy clock
> > > > > > > cycles), but there might be others aswell. So by removing this functionality
> > > > > > > above use case will brake, this since those test will not be reliable.
> > > > > > > Furthermore, since users tend to be creative it is not possible to know if
> > > > > > > there are other use cases that will be affected. This means that in case [1]
> > > > > > > needs to be followed the safe path is to add functionality instead of removing.
> > > > > > > Luckily it also easier in this case, see below. > > > > > >
> > > > > > I understand there might be users other than U-Boot/Linux that use an
> > > > > > odd number of dummy bits (not multiple of 8). If your concern was
> > > > > > about model behavior changes, sure I can update > > > > > > qemu/docs/system/deprecated.rst to mention that some flashes in the
> > > > > > m25p80 model now implement dummy cycles as bytes. > > > > >
> > > > > Yes, something like that. My concern is that since this functionality has been
> > > > > in tree for while, users have found known or unknown features that got
> > > > > introduced by it. By removing the functionality (and the known/uknown features)
> > > > > we are riscing to brake our user's use cases (currently I'm aware of one
> > > > > feature/use case but it is not unlikely that there are more).
[1] states that
> > > > > "In general features are intended to be supported indefinitely once introduced
> > > > > into QEMU", to me that makes very much sense because the opposite would mean
> > > > > that we were not reliable. So in case [1] needs to be honored it looks to be
> > > > > safer to add functionality instead of removing (and riscing the removal of use
> > > > > cases/features). Luckily I still believe in this case that it will be easier to
> > > > > go forward (even if I also agree on what you are saying below about what I
> > > > > proposed).
> > > > >
> > > >
> > > > Even if the implementation is buggy and we need to keep the buggy
> > > > implementation forever? I think that's why
> > > > qemu/docs/system/deprecated.rst was created for deprecating such > > > > feature.
> > >
> > > With the RFC I posted all commands in m25p80 are working for both the case 1
> > > controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI).
> > > Because of this, I, with all respect, will have to disagree that this is buggy.
> >
> > Well, the existing m25p80 implementation that uses dummy cycle > > accuracy for those flashes prevents all SPI controllers that use tx > > fifo to work with those flashes. Hence it is buggy.
> >
> > >
> > > >
> > > > > >
> > > > > > > >
> > > > > > > > > don't think it is fair to call them 'seriously broken' (and else we should
> > > > > > > > > probably let the maintainers know about it). Most likely the lack of support
> > > > > > > >
> > > > > > > > I called it "seriously broken" because current implementation only
> > > > > > > > considered one type of SPI controllers while completely ignoring the
> > > > > > > > other type.
> > > > > > >
> > > > > > > If we change view and see this from the perspective of m25p80, it models the
> > > > > > > commands a certain way and provides an API that the SPI controllers need to
> > > > > > > implement for interacting with it. It is true that there are SPI controllers
> > > > > > > referred to above that do not support the portion of that API that corresponds
> > > > > > > to commands with dummy clock cycles, but I don't think it is true that this is
> > > > > > > broken since there is also one SPI controller that has a working implementation
> > > > > > > of m25p80's full API also when transfering through a tx fifo (use case 1). But
> > > > > > > as mentioned above, by doing a minor extension and improvement to m25p80's API
> > > > > > > and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1]
> > > > > > > will still be honored as in the same time making it possible to have full
> > > > > > > support for the API in the SPI controllers that currently do not (please reread
> > > > > > > the proposal in my previous reply that attempts to do this). I myself see this
> > > > > > > as win/win situation, also because no controller should need modifications.
> > > > > > >
> > > > > >
> > > > > > I am afraid your proposal does not work. Your proposed new device
> > > > > > property 'model_dummy_bytes' to select to convert the accurate dummy
> > > > > > clock cycle count to dummy bytes inside m25p80, is hard to justify as
> > > > > > a property to the flash itself, as the behavior is tightly coupled to
> > > > > > how the SPI controller works.
> > > > >
> > > > > I agree on above. I decided though that instead of posting sample code in here
> > > > > I'll post an RFC with hopefully an improved proposal. I'll cc you. About below,
> > > > > Xilinx ZynqMP GQSPI should not need any modication in a first step.
> > > > >
> > > >
> > > > Wait, (see below)
> > > >
> > > > > >
> > > > > > Please take a look at the Xilinx GQSPI controller, which supports both
> > > > > > use cases, that the dummy cycles can be transferred via tx fifo, or
> > > > > > generated by the controller automatically. Please read the example
> > > > > > given in:
> > > > > >
> > > > > >     table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
> > > > > > Command (EBh)
> > > > > >
> > > > > > in
https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf
> > > > > >
> > > > > > If you choose to set the m25p80 device property 'model_dummy_bytes' to
> > > > > > true when working with the Xilinx GQSPI controller, you are bound to
> > > > > > only allow guest software to use tx fifo to transfer the dummy cycles,
> > > > > > and this is wrong.
> > > > > >
> > > >
> > > > You missed this part. I looked at your RFC, and as I mentioned above
> > > > your proposal cannot support the complicated controller like Xilinx
> > > > GQSPI. Please read the example of table 24-22. With your RFC, you
> > > > mandate guest software's GQSPI driver to only use hardware dummy cycle
> > > > generation, which is wrong.
> > > >
> > >
> > > First, thank you very much for looking into the RFC series, very much
> > > appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2
> > > locations in the file, in 1 location the transfer referred to above is done, in
> > > another location the transfer through the txfifo is done. The location where
> > > transfer referred to above is done will not need any modifications (and will
> > > thus work equally well as it does currently).
> >
> > Please explain this a little bit. How does your RFC series handle > > cases as described in table 24-22, where the 6 dummy cycles are split
> > into 2 transfers, with one transfer using tx fifo, and the other one > > using hardware dummy cycle generation?
>
> Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

Quoted 6 lines

Francisco really knows this stuff better than me....
I would tend to agree that it's unfortunate to model things in cycles, if we could abstract things at a higher level that would be nice. Without breaking existing use-cases.
Francisco, is it impossible to bring up the abstraction level to bytes and keep existing use-cases?

Great question, I'm leaning on that it shouldn't be impossible to be honest (but I haven't been able to try anything yet though).

Best regards,
Francisco Iglesias

Quoted 5 lines

We have a bunch of test-cases, We'll publish some of them in source code, others we can't publish since they use proprietary SW we're not allowed to publish at all, but we can run tests and Ack if things work. Best regards,
Edgar

author
#30 lore raw
In reply to Bin Meng
Quoted 227 lines · Bin Meng
Quoted 211 lines · Francisco Iglesias

Hi Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

Quoted 17 lines · Bin Meng

From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up
bytes are expected to be received after it receives a command. For
example, depending on the address mode, either 3-byte address or
4-byte address is needed.

For fast read family commands, some dummy cycles are required after
sending the address bytes, and the dummy cycles need to be counted
in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte.
It is not in bit, or cycle. However for some reason the model has
been using the number of dummy cycles for s->needed_bytes. The right
approach is to convert the number of dummy cycles to bytes based on
the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad
I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was considered but not chosen by the developers due to it is inaccuracy (it wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8, meaning that if the controller is wrongly programmed to generate 7 the error wouldn't be caught and the controller will still be considered "correct"). Now that we have this detail in the implementation I'm in favor of keeping it, this also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 69 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 6 lines
Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Quoted 5 lines

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

Hello maintainers,

We apparently missed the 6.0 window to address this mess of the m25p80 model. Please provide your inputs on this before I start working on the v2.

Regards,
Bin

#31 lore raw
Quoted 229 lines · Bin Meng
Quoted 227 lines · Bin Meng
Quoted 211 lines · Francisco Iglesias

Hi Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

Quoted 30 lines · Francisco Iglesias

Hi Bin,

On [2021 Jan 14] Thu 23:08:53, Bin Meng wrote:
From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up
bytes are expected to be received after it receives a command. For
example, depending on the address mode, either 3-byte address or
4-byte address is needed.

For fast read family commands, some dummy cycles are required after
sending the address bytes, and the dummy cycles need to be counted
in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte.
It is not in bit, or cycle. However for some reason the model has
been using the number of dummy cycles for s->needed_bytes. The right
approach is to convert the number of dummy cycles to bytes based on
the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad
I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was
considered but not chosen by the developers due to it is inaccuracy (it
wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8,
meaning that if the controller is wrongly programmed to generate 7 the error
wouldn't be caught and the controller will still be considered "correct"). Now
that we have this detail in the implementation I'm in favor of keeping it, this
also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here was ever considered, otherwise all SPI controller models supporting software generation should have been found out seriously broken long time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 69 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 6 lines
Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Quoted 5 lines
Quoted 5 lines

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

I agree with Edgar, in that Francisco and Bin know this better than me and that modelling things in cycles is a pain.

As Bin points out it seems like currently we should be modelling bytes (from the variable name) so it makes sense to keep it in bytes. I would be in favour of this series in that case. Do we know what use cases this will break? I know it's hard to answer but I don't think there are too many SSI users in QEMU so it might not be too hard to test most of the possible use cases.

Alistair

Quoted 10 lines

Hello maintainers,

We apparently missed the 6.0 window to address this mess of the m25p80 model. Please provide your inputs on this before I start working on the v2.

Regards,
Bin

#32 lore raw
Quoted 230 lines · Alistair Francis
Quoted 229 lines · Bin Meng
Quoted 227 lines · Bin Meng
Quoted 211 lines · Francisco Iglesias

Hi Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 39 lines · Francisco Iglesias

Hi Bin,

Quoted 35 lines · Bin Meng

Hi Francisco,

On Fri, Jan 15, 2021 at 2:13 AM Francisco Iglesias
<[email protected]> wrote:

Hi Bin,

On [2021 Jan 14] Thu 23:08:53, Bin Meng wrote:
From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up
bytes are expected to be received after it receives a command. For
example, depending on the address mode, either 3-byte address or
4-byte address is needed.

For fast read family commands, some dummy cycles are required after
sending the address bytes, and the dummy cycles need to be counted
in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte.
It is not in bit, or cycle. However for some reason the model has
been using the number of dummy cycles for s->needed_bytes. The right
approach is to convert the number of dummy cycles to bytes based on
the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad
I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was
considered but not chosen by the developers due to it is inaccuracy (it
wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8,
meaning that if the controller is wrongly programmed to generate 7 the error
wouldn't be caught and the controller will still be considered "correct"). Now
that we have this detail in the implementation I'm in favor of keeping it, this
also because the detail is already in use for catching exactly above error.

Quoted 8 lines
Quoted 4 lines

I found no clue from the commit message that my proposed solution here
was ever considered, otherwise all SPI controller models supporting
software generation should have been found out seriously broken long
time ago!

The controllers you are referring to might lack support for commands requiring dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 69 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

don't think it is fair to call them 'seriously broken' (and else we should probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 6 lines
Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Quoted 8 lines
Quoted 5 lines
Quoted 5 lines

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

I agree with Edgar, in that Francisco and Bin know this better than me and that modelling things in cycles is a pain.

Hi Alistair,

Quoted 7 lines

As Bin points out it seems like currently we should be modelling bytes (from the variable name) so it makes sense to keep it in bytes. I would be in favour of this series in that case. Do we know what use cases this will break? I know it's hard to answer but I don't think there are too many SSI users in QEMU so it might not be too hard to test most of the possible use cases.

The use case I'm aware of is regression testing of drivers. Ex: if a driver is using 10 dummy clock cycles with the commands and a patch accidentaly changes the driver to use 11 dummy clock cycles QEMU currently finds the problem, that won't be possible with this series. It's difficult to say but it is not impossible there are other use cases also.

More importantly IMO though is that the current use cases can be keept while still providing support for commands with dummy clock cycles into the QEMU SPI controllers lacking at the moment.

(If I recall correctly this series might also have another issue regarding the GQSPI SPI mode configuration, with that it is possible transmit 8 dummy clock cycles as 1 data byte, 2 data bytes or 4 data bytes, so I think some form of calculation might be needed inside m25p80).

Best regards,
Francisco

Quoted 13 lines

Alistair

Quoted 10 lines

Hello maintainers,

We apparently missed the 6.0 window to address this mess of the m25p80 model. Please provide your inputs on this before I start working on the v2.

Regards,
Bin

#33 lore raw

Hello,

Quoted 250 lines · Francisco Iglesias
Quoted 230 lines · Alistair Francis
Quoted 229 lines · Bin Meng
Quoted 227 lines · Bin Meng
Quoted 211 lines · Francisco Iglesias

Hi Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 52 lines · Bin Meng

Hi Francisco,

Quoted 47 lines · Francisco Iglesias

Hi Bin,

On [2021 Jan 15] Fri 10:07:52, Bin Meng wrote:
Hi Francisco,

On Fri, Jan 15, 2021 at 2:13 AM Francisco Iglesias
<[email protected]> wrote:

Hi Bin,

On [2021 Jan 14] Thu 23:08:53, Bin Meng wrote:
From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up
bytes are expected to be received after it receives a command. For
example, depending on the address mode, either 3-byte address or
4-byte address is needed.

For fast read family commands, some dummy cycles are required after
sending the address bytes, and the dummy cycles need to be counted
in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte.
It is not in bit, or cycle. However for some reason the model has
been using the number of dummy cycles for s->needed_bytes. The right
approach is to convert the number of dummy cycles to bytes based on
the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad
I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was
considered but not chosen by the developers due to it is inaccuracy (it
wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8,
meaning that if the controller is wrongly programmed to generate 7 the error
wouldn't be caught and the controller will still be considered "correct"). Now
that we have this detail in the implementation I'm in favor of keeping it, this
also because the detail is already in use for catching exactly above error.

I found no clue from the commit message that my proposed solution here
was ever considered, otherwise all SPI controller models supporting
software generation should have been found out seriously broken long
time ago!

The controllers you are referring to might lack support for commands requiring
dummy clock cycles but I really hope they work with the other commands? If so I

Quoted 8 lines

I am not sure why you view dummy clock cycles as something special that needs some special support from the SPI controller. For the case 1 controller, it's nothing special from the controller perspective, just like sending out a command, or address bytes, or data. The controller just shifts data bit by bit from its tx fifo and that's it. In the Xilinx GQSPI controller case, the dummy cycles can either be sent via a regular data (the case 1 controller) in the tx fifo, or automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 69 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines
Quoted 6 lines

don't think it is fair to call them 'seriously broken' (and else we should
probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only considered one type of SPI controllers while completely ignoring the other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 6 lines
Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Quoted 8 lines
Quoted 5 lines
Quoted 5 lines

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

I agree with Edgar, in that Francisco and Bin know this better than me and that modelling things in cycles is a pain.

Hi Alistair,

Quoted 7 lines

As Bin points out it seems like currently we should be modelling bytes (from the variable name) so it makes sense to keep it in bytes. I would be in favour of this series in that case. Do we know what use cases this will break? I know it's hard to answer but I don't think there are too many SSI users in QEMU so it might not be too hard to test most of the possible use cases.

Quoted 5 lines

The use case I'm aware of is regression testing of drivers. Ex: if a driver is using 10 dummy clock cycles with the commands and a patch accidentaly changes the driver to use 11 dummy clock cycles QEMU currently finds the problem, that won't be possible with this series. It's difficult to say but it is not impossible there are other use cases also.

It was breaking the Aspeed machines :

https://lore.kernel.org/qemu-devel/[email protected]/

QEMU 6.1 should have acceptance tests that will help in detecting regressions in this area.

Thanks,

C.

Quoted 27 lines

More importantly IMO though is that the current use cases can be keept while still providing support for commands with dummy clock cycles into the QEMU SPI controllers lacking at the moment.

(If I recall correctly this series might also have another issue regarding the GQSPI SPI mode configuration, with that it is possible transmit 8 dummy clock cycles as 1 data byte, 2 data bytes or 4 data bytes, so I think some form of calculation might be needed inside m25p80).

Best regards,
Francisco

Quoted 13 lines

Alistair

Quoted 10 lines

Hello maintainers,

We apparently missed the 6.0 window to address this mess of the m25p80 model. Please provide your inputs on this before I start working on the v2.

Regards,
Bin

author
#34 lore raw

Hi Cédric,

Quoted 261 lines · Cédric Le Goater

Hello,

Quoted 250 lines · Francisco Iglesias
Quoted 230 lines · Alistair Francis
Quoted 229 lines · Bin Meng
Quoted 227 lines · Bin Meng
Quoted 211 lines · Francisco Iglesias

Hi Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 71 lines · Francisco Iglesias

Hi Bin,

Quoted 60 lines · Bin Meng

Hi Francisco,

On Fri, Jan 15, 2021 at 8:26 PM Francisco Iglesias
<[email protected]> wrote:

Hi Bin,

On [2021 Jan 15] Fri 10:07:52, Bin Meng wrote:
Hi Francisco,

On Fri, Jan 15, 2021 at 2:13 AM Francisco Iglesias
<[email protected]> wrote:

Hi Bin,

On [2021 Jan 14] Thu 23:08:53, Bin Meng wrote:
From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up
bytes are expected to be received after it receives a command. For
example, depending on the address mode, either 3-byte address or
4-byte address is needed.

For fast read family commands, some dummy cycles are required after
sending the address bytes, and the dummy cycles need to be counted
in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte.
It is not in bit, or cycle. However for some reason the model has
been using the number of dummy cycles for s->needed_bytes. The right
approach is to convert the number of dummy cycles to bytes based on
the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad
I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was
considered but not chosen by the developers due to it is inaccuracy (it
wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8,
meaning that if the controller is wrongly programmed to generate 7 the error
wouldn't be caught and the controller will still be considered "correct"). Now
that we have this detail in the implementation I'm in favor of keeping it, this
also because the detail is already in use for catching exactly above error.

I found no clue from the commit message that my proposed solution here
was ever considered, otherwise all SPI controller models supporting
software generation should have been found out seriously broken long
time ago!

The controllers you are referring to might lack support for commands requiring
dummy clock cycles but I really hope they work with the other commands? If so I

I am not sure why you view dummy clock cycles as something special
that needs some special support from the SPI controller. For the case
1 controller, it's nothing special from the controller perspective,
just like sending out a command, or address bytes, or data. The
controller just shifts data bit by bit from its tx fifo and that's it.
In the Xilinx GQSPI controller case, the dummy cycles can either be
sent via a regular data (the case 1 controller) in the tx fifo, or
automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also need to keep in mind that QEMU models HW, and any binary that runs on a HW board supported in QEMU should ideally run on that board inside QEMU aswell (this can be a bare metal application equaly well as a modified u-boot/Linux using SPI commands with a non multiple of 8 number of dummy clock cycles).

Quoted 12 lines

Once functionality has been introduced into QEMU it is not easy to know which intentional or untentional features provided by the functionality are being used by users. One of the (perhaps not well known) features I'm aware of that is in use and is provided by the accurate dummy clock cycle modeling inside m25p80 is the be ability to test drivers accurately regarding the dummy clock cycles (even when using commands with a non-multiple of 8 number of dummy clock cycles), but there might be others aswell. So by removing this functionality above use case will brake, this since those test will not be reliable. Furthermore, since users tend to be creative it is not possible to know if there are other use cases that will be affected. This means that in case [1] needs to be followed the safe path is to add functionality instead of removing. Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 69 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 11 lines
Quoted 10 lines

don't think it is fair to call them 'seriously broken' (and else we should
probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only
considered one type of SPI controllers while completely ignoring the
other type.

Quoted 14 lines

If we change view and see this from the perspective of m25p80, it models the commands a certain way and provides an API that the SPI controllers need to implement for interacting with it. It is true that there are SPI controllers referred to above that do not support the portion of that API that corresponds to commands with dummy clock cycles, but I don't think it is true that this is broken since there is also one SPI controller that has a working implementation of m25p80's full API also when transfering through a tx fifo (use case 1). But as mentioned above, by doing a minor extension and improvement to m25p80's API and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1] will still be honored as in the same time making it possible to have full support for the API in the SPI controllers that currently do not (please reread the proposal in my previous reply that attempts to do this). I myself see this as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 6 lines
Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Quoted 8 lines
Quoted 5 lines
Quoted 5 lines

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

I agree with Edgar, in that Francisco and Bin know this better than me and that modelling things in cycles is a pain.

Hi Alistair,

Quoted 7 lines

As Bin points out it seems like currently we should be modelling bytes (from the variable name) so it makes sense to keep it in bytes. I would be in favour of this series in that case. Do we know what use cases this will break? I know it's hard to answer but I don't think there are too many SSI users in QEMU so it might not be too hard to test most of the possible use cases.

Quoted 5 lines

The use case I'm aware of is regression testing of drivers. Ex: if a driver is using 10 dummy clock cycles with the commands and a patch accidentaly changes the driver to use 11 dummy clock cycles QEMU currently finds the problem, that won't be possible with this series. It's difficult to say but it is not impossible there are other use cases also.

Quoted 3 lines

It was breaking the Aspeed machines :

https://lore.kernel.org/qemu-devel/[email protected]/

Yes, as I mentioned in the series the modification was based on a pure guess from existing QEMU codes as I could not find a datasheet of the Aspeed SPI controller on the internet. Do you know if this is publicly available?

Quoted 4 lines

QEMU 6.1 should have acceptance tests that will help in detecting regressions in this area.

Regards,
Bin

#35 lore raw
Quoted 264 lines · Bin Meng

Hi Cédric,

Quoted 261 lines · Cédric Le Goater

Hello,

Quoted 255 lines · Francisco Iglesias
Quoted 230 lines · Alistair Francis
Quoted 229 lines · Bin Meng
Quoted 227 lines · Bin Meng
Quoted 211 lines · Francisco Iglesias

Hi Bin,

Quoted 207 lines · Bin Meng

Hi Francisco,

Quoted 120 lines · Francisco Iglesias

Dear Bin,

Quoted 116 lines · Bin Meng

Hi Francisco,

Quoted 111 lines · Francisco Iglesias

Hi Bin,

Quoted 88 lines · Bin Meng

Hi Francisco,

Quoted 83 lines · Francisco Iglesias

Hi Bin,

On [2021 Jan 15] Fri 22:38:18, Bin Meng wrote:
Hi Francisco,

On Fri, Jan 15, 2021 at 8:26 PM Francisco Iglesias
<[email protected]> wrote:

Hi Bin,

On [2021 Jan 15] Fri 10:07:52, Bin Meng wrote:
Hi Francisco,

On Fri, Jan 15, 2021 at 2:13 AM Francisco Iglesias
<[email protected]> wrote:

Hi Bin,

On [2021 Jan 14] Thu 23:08:53, Bin Meng wrote:
From: Bin Meng <[email protected]>

The m25p80 model uses s->needed_bytes to indicate how many follow-up
bytes are expected to be received after it receives a command. For
example, depending on the address mode, either 3-byte address or
4-byte address is needed.

For fast read family commands, some dummy cycles are required after
sending the address bytes, and the dummy cycles need to be counted
in s->needed_bytes. This is where the mess began.

As the variable name (needed_bytes) indicates, the unit is in byte.
It is not in bit, or cycle. However for some reason the model has
been using the number of dummy cycles for s->needed_bytes. The right
approach is to convert the number of dummy cycles to bytes based on
the SPI protocol, for example, 6 dummy cycles for the Fast Read Quad
I/O (EBh) should be converted to 3 bytes per the formula (6 * 4 / 8).

While not being the original implementor I must assume that above solution was
considered but not chosen by the developers due to it is inaccuracy (it
wouldn't be possible to model exacly 6 dummy cycles, only a multiple of 8,
meaning that if the controller is wrongly programmed to generate 7 the error
wouldn't be caught and the controller will still be considered "correct"). Now
that we have this detail in the implementation I'm in favor of keeping it, this
also because the detail is already in use for catching exactly above error.

I found no clue from the commit message that my proposed solution here
was ever considered, otherwise all SPI controller models supporting
software generation should have been found out seriously broken long
time ago!

The controllers you are referring to might lack support for commands requiring
dummy clock cycles but I really hope they work with the other commands? If so I

I am not sure why you view dummy clock cycles as something special
that needs some special support from the SPI controller. For the case
1 controller, it's nothing special from the controller perspective,
just like sending out a command, or address bytes, or data. The
controller just shifts data bit by bit from its tx fifo and that's it.
In the Xilinx GQSPI controller case, the dummy cycles can either be
sent via a regular data (the case 1 controller) in the tx fifo, or
automatically generated (case 2 controller) by the hardware.

Ok, I'll try to explain my view point a little differently. For that we also
need to keep in mind that QEMU models HW, and any binary that runs on a HW
board supported in QEMU should ideally run on that board inside QEMU aswell
(this can be a bare metal application equaly well as a modified u-boot/Linux
using SPI commands with a non multiple of 8 number of dummy clock cycles).

Once functionality has been introduced into QEMU it is not easy to know which
intentional or untentional features provided by the functionality are being
used by users. One of the (perhaps not well known) features I'm aware of that
is in use and is provided by the accurate dummy clock cycle modeling inside
m25p80 is the be ability to test drivers accurately regarding the dummy clock
cycles (even when using commands with a non-multiple of 8 number of dummy clock
cycles), but there might be others aswell. So by removing this functionality
above use case will brake, this since those test will not be reliable.
Furthermore, since users tend to be creative it is not possible to know if
there are other use cases that will be affected. This means that in case [1]
needs to be followed the safe path is to add functionality instead of removing.
Luckily it also easier in this case, see below.

Quoted 5 lines

I understand there might be users other than U-Boot/Linux that use an odd number of dummy bits (not multiple of 8). If your concern was about model behavior changes, sure I can update qemu/docs/system/deprecated.rst to mention that some flashes in the m25p80 model now implement dummy cycles as bytes.

Yes, something like that. My concern is that since this functionality has been in tree for while, users have found known or unknown features that got introduced by it. By removing the functionality (and the known/uknown features) we are riscing to brake our user's use cases (currently I'm aware of one feature/use case but it is not unlikely that there are more). [1] states that "In general features are intended to be supported indefinitely once introduced into QEMU", to me that makes very much sense because the opposite would mean that we were not reliable. So in case [1] needs to be honored it looks to be safer to add functionality instead of removing (and riscing the removal of use cases/features). Luckily I still believe in this case that it will be easier to go forward (even if I also agree on what you are saying below about what I proposed).

Quoted 8 lines
Quoted 4 lines

Even if the implementation is buggy and we need to keep the buggy implementation forever? I think that's why qemu/docs/system/deprecated.rst was created for deprecating such feature.

With the RFC I posted all commands in m25p80 are working for both the case 1 controller (using a txfifo) and the case 2 controller (no txfifo, as GQSPI). Because of this, I, with all respect, will have to disagree that this is buggy.

Well, the existing m25p80 implementation that uses dummy cycle accuracy for those flashes prevents all SPI controllers that use tx fifo to work with those flashes. Hence it is buggy.

Quoted 69 lines
Quoted 56 lines
Quoted 32 lines
Quoted 26 lines
Quoted 25 lines

don't think it is fair to call them 'seriously broken' (and else we should
probably let the maintainers know about it). Most likely the lack of support

I called it "seriously broken" because current implementation only
considered one type of SPI controllers while completely ignoring the
other type.

If we change view and see this from the perspective of m25p80, it models the
commands a certain way and provides an API that the SPI controllers need to
implement for interacting with it. It is true that there are SPI controllers
referred to above that do not support the portion of that API that corresponds
to commands with dummy clock cycles, but I don't think it is true that this is
broken since there is also one SPI controller that has a working implementation
of m25p80's full API also when transfering through a tx fifo (use case 1). But
as mentioned above, by doing a minor extension and improvement to m25p80's API
and allow for toggling the accuracy from dummy clock cycles to dummy bytes [1]
will still be honored as in the same time making it possible to have full
support for the API in the SPI controllers that currently do not (please reread
the proposal in my previous reply that attempts to do this). I myself see this
as win/win situation, also because no controller should need modifications.

Quoted 5 lines

I am afraid your proposal does not work. Your proposed new device property 'model_dummy_bytes' to select to convert the accurate dummy clock cycle count to dummy bytes inside m25p80, is hard to justify as a property to the flash itself, as the behavior is tightly coupled to how the SPI controller works.

Quoted 4 lines

I agree on above. I decided though that instead of posting sample code in here I'll post an RFC with hopefully an improved proposal. I'll cc you. About below, Xilinx ZynqMP GQSPI should not need any modication in a first step.

Wait, (see below)

Quoted 9 lines
Quoted 9 lines

Please take a look at the Xilinx GQSPI controller, which supports both use cases, that the dummy cycles can be transferred via tx fifo, or generated by the controller automatically. Please read the example given in:

table 24‐22, an example of Generic FIFO Contents for Quad I/O Read
Command (EBh)

Quoted 7 lines
Quoted 7 lines

in https://www.xilinx.com/support/documentation/user_guides/ug1085-zynq-ultrascale-trm.pdf

If you choose to set the m25p80 device property 'model_dummy_bytes' to true when working with the Xilinx GQSPI controller, you are bound to only allow guest software to use tx fifo to transfer the dummy cycles, and this is wrong.

Quoted 6 lines

You missed this part. I looked at your RFC, and as I mentioned above your proposal cannot support the complicated controller like Xilinx GQSPI. Please read the example of table 24-22. With your RFC, you mandate guest software's GQSPI driver to only use hardware dummy cycle generation, which is wrong.

First, thank you very much for looking into the RFC series, very much appreciated. Secondly, about above, the GQSPI model in QEMU transfers from 2 locations in the file, in 1 location the transfer referred to above is done, in another location the transfer through the txfifo is done. The location where transfer referred to above is done will not need any modifications (and will thus work equally well as it does currently).

Quoted 6 lines
Quoted 4 lines

Please explain this a little bit. How does your RFC series handle cases as described in table 24-22, where the 6 dummy cycles are split into 2 transfers, with one transfer using tx fifo, and the other one using hardware dummy cycle generation?

Sorry, I missunderstod. You are right, that won't work.

+Edgar E. Iglesias

So it looks by far the only way to implement dummy cycles correctly to work with all SPI controller models is what I proposed here in this patch series.

Quoted 8 lines
Quoted 5 lines
Quoted 5 lines

Maintainers are quite silent, so I would like to hear your thoughts.

@Alistair Francis @Philippe Mathieu-Daudé @Peter Maydell would you please share your thoughts since you are the one who reviewed the existing dummy implementation (based on commits history)

I agree with Edgar, in that Francisco and Bin know this better than me and that modelling things in cycles is a pain.

Hi Alistair,

Quoted 7 lines

As Bin points out it seems like currently we should be modelling bytes (from the variable name) so it makes sense to keep it in bytes. I would be in favour of this series in that case. Do we know what use cases this will break? I know it's hard to answer but I don't think there are too many SSI users in QEMU so it might not be too hard to test most of the possible use cases.

The use case I'm aware of is regression testing of drivers. Ex: if a driver is using 10 dummy clock cycles with the commands and a patch accidentaly changes the driver to use 11 dummy clock cycles QEMU currently finds the problem, that won't be possible with this series. It's difficult to say but it is not impossible there are other use cases also.

Quoted 8 lines
Quoted 3 lines

It was breaking the Aspeed machines :

https://lore.kernel.org/qemu-devel/[email protected]/

Yes, as I mentioned in the series the modification was based on a pure guess from existing QEMU codes as I could not find a datasheet of the Aspeed SPI controller on the internet. Do you know if this is publicly available?

It is not but much of the register bitfields are described in the code. I should be able to help you in making this work.

Thanks,

C.

Quoted 7 lines
Quoted 3 lines

QEMU 6.1 should have acceptance tests that will help in detecting regressions in this area.

Regards,
Bin