Skip to contentSkip to conversation

sdhci: fix SDMA boundary bug

PATCH Superseded Tao Ding 2 participants 5 messages Jul 1 – Jul 6

Subsystems (2)

v3 of 7 Superseded Latest is v7 →

lore
Share & tools
lore

Tao Ding

author
#0 lore raw

v3 cover sdhci: fix SDMA boundary bug

Hi all,

This series fixes an SDMA multi-block transfer bug in the generic SDHCI controller.

v2 of this series was posted here:
https://lore.kernel.org/qemu-devel/[email protected]/

Thanks

Changes in v2->v3:
* The code of v2 will cause the register protected by TRANSFERRING_DATA to be modified incorrectly. v3 avoid this problem.

Tao Ding (2):
hw/sd: sdhci: Fix SMDA boundary bug
qtest: add xilinx-zynq SDHCI sdma test

 hw/sd/sdhci.c                        |   9 +-
 include/hw/sd/sdhci.h                |   2 +
 tests/qtest/meson.build              |   1 +
 tests/qtest/xilinx-zynq-sdhci-test.c | 178 +++++++++++++++++++++++++++
 4 files changed, 189 insertions(+), 1 deletion(-)
 create mode 100644 tests/qtest/xilinx-zynq-sdhci-test.c

base-commit: 608ad9307d5b4797d2254ee3a4bad2bc8c8bcbae

Signature · 1 line
2.43.0
author
#1 lore raw

PATCH v3 1/2 hw/sd: sdhci: Fix SDMA boundary bug

During SDMA transfers, the controller raises an interrupt at buffer boundaries to request a system address update. The host driver will rewrite the SDHC_SYSAD register. (according to PartA2_SD_Host_Controller_Simplified_Specification_Ver2.00.pdf section 2.2.1) However, the current code will ignore write operations to SDHC_SYSAD.

To fix this bug, when SDMA encounters a boundary, a state is set to record it.
In this state, SDHC_SYSAD can be written. In other cases, it is still protected by TRANSFERRING_DATA.

---
The reviewer pointed out an issue in patch v2 and suggested a more appropriate way of modification. In this patch, a new state has been added to indicate encountering a boundary.
No additional state has been added to the existing stopped_state,
which seems to be specifically for gap requests (as can be seen from some if-else statements in the code)

This bug may be encountered when using uboot with SDMA functionality, and can be reproduced by following the steps below. Uboot test:
1. Prepare zImage, uboot and rootfs.cpio.gz, make sure uboot config (MMC_SDHCI_SDMA=y).
for more detail, can obtained "https://lore.kernel.org/qemu-devel/[email protected]/T/#mcdf231e6bd762903f8393bf93253cbcb12ececcf"
2. ./qemu-system-aarch64 -M xilinx-zynq-a9 -machine boot-mode=sd -m 1024 -display none -serial null -serial stdio -monitor none \
-device loader,file=u-boot-dtb.bin,addr=0x04000000,cpu-num=0 \
-drive file=zynq-sd.img,format=raw,if=sd -D qemu.log

After applying the fix, success load sd card:

U-Boot 2026.07-rc4-g1e80ee41441c (Jun 15 2026 - 19:05:05 +0800)

Model: Xilinx ZC702 board
DRAM:  ECC disabled 1 GiB
Core:  33 devices, 21 uclasses, devicetree: board
Flash: 0 Bytes
NAND:  0 MiB
MMC:   mmc@e0100000: 0

Loading Environment from FAT... *** Error - No Valid Environment Area found *** Warning - bad env area, using default environment

In:    serial@e0001000
Out:   serial@e0001000
Err:   serial@e0001000

Net:
ZYNQ GEM: e000b000, mdio bus e000b000, phyaddr 7, interface rgmii-id

Warning: ethernet@e000b000 (eth0) using random MAC address - 4a:63:5a:a4:d3:6b
eth0: ethernet@e000b000
Hit any key to stop autoboot: 0
Zynq> fatload mmc 0 ${kernel_addr_r} zImage
10838528 bytes read in 4316 ms (2.4 MiB/s)
Zynq> fatload mmc 0 ${ramdisk_addr_r} rootfs-arm32.cpio.gz
406531 bytes read in 199 ms (1.9 MiB/s)
Zynq> setenv ramdisk_size ${filesize}
Zynq> fatload mmc 0 ${fdt_addr_r} zynq-zc702.dtb
15754 bytes read in 21 ms (732.4 KiB/s)
Zynq> setenv bootargs console=ttyPS0,115200 earlycon ignore_loglevel rdinit=/init Zynq> bootz ${kernel_addr_r} ${ramdisk_addr_r}:${ramdisk_size} ${fdt_addr_r} Kernel image @ 0x2000000 [ 0x000000 - 0xa56200 ]
## Flattened Device Tree blob at 01f00000
Booting using the fdt blob at 0x1f00000
Working FDT set to 1f00000
Loading Ramdisk to 2ff9c000, end 2ffff403 ... OK
Loading Device Tree to 2ff95000, end 2ff9bd89 ... OK
Working FDT set to 2ff95000

Starting kernel ...

 hw/sd/sdhci.c         | 9 ++++++++-
 include/hw/sd/sdhci.h | 2 ++
 2 files changed, 10 insertions(+), 1 deletion(-)
hw/sd/sdhci.c +8 -1
index c86dfa281f..ecd286f2d7 100644
@@ -307,6 +307,7 @@ static void sdhci_reset(SDHCIState *s)
     s->data_count = 0;
     s->stopped_state = sdhc_not_stopped;
     s->pending_insert_state = false;
+    s->sdma_boundary_paused = false;
     if (object_dynamic_cast(OBJECT(s), TYPE_FSL_ESDHC_BE) ||
             object_dynamic_cast(OBJECT(s), TYPE_FSL_ESDHC_LE)) {
         s->norintstsen = 0x013f;
@@ -620,6 +621,7 @@ static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s)
     }
 
     s->prnsts |= SDHC_DATA_INHIBIT | SDHC_DAT_LINE_ACTIVE;
+    s->sdma_boundary_paused = false;
     if (s->trnmod & SDHC_TRNS_READ) {
         s->prnsts |= SDHC_DOING_READ;
         while (s->blkcnt) {
@@ -644,6 +646,7 @@ static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s)
                 s->data_count = 0;
             }
             if (page_aligned && boundary_count == 0) {
+                s->sdma_boundary_paused = true;
                 break;
             }
         }
@@ -669,6 +672,7 @@ static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s)
                 }
             }
             if (page_aligned && boundary_count == 0) {
+                s->sdma_boundary_paused = true;
                 break;
             }
         }
@@ -679,6 +683,7 @@ static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s)
     }
 
     if (s->blkcnt == 0) {
+        s->sdma_boundary_paused = false;
         sdhci_end_transfer(s);
     } else {
         sdhci_update_irq(s);
@@ -1164,6 +1169,7 @@ static inline void sdhci_reset_write(SDHCIState *s, uint8_t value)
                 SDHC_DATA_INHIBIT | SDHC_DAT_LINE_ACTIVE);
         s->blkgap &= ~(SDHC_STOP_AT_GAP_REQ | SDHC_CONTINUE_REQ);
         s->stopped_state = sdhc_not_stopped;
+        s->sdma_boundary_paused = false;
         s->norintsts &= ~(SDHC_NIS_WBUFRDY | SDHC_NIS_RBUFRDY |
                 SDHC_NIS_DMA | SDHC_NIS_TRSCMP | SDHC_NIS_BLKGAP);
         break;
@@ -1185,7 +1191,7 @@ sdhci_write(void *opaque, hwaddr offset, uint64_t val, unsigned size)
 
     switch (offset & ~0x3) {
     case SDHC_SYSAD:
-        if (!TRANSFERRING_DATA(s->prnsts)) {
+        if (!TRANSFERRING_DATA(s->prnsts) || s->sdma_boundary_paused) {
             s->sdmasysad = (s->sdmasysad & mask) | value;
             MASKED_WRITE(s->sdmasysad, mask, value);
             /* Writing to last byte of sdmasysad might trigger transfer */
@@ -1508,6 +1514,7 @@ const VMStateDescription sdhci_vmstate = {
         VMSTATE_VBUFFER_UINT32(fifo_buffer, SDHCIState, 1, NULL, buf_maxsz),
         VMSTATE_TIMER_PTR(insert_timer, SDHCIState),
         VMSTATE_TIMER_PTR(transfer_timer, SDHCIState),
+        VMSTATE_BOOL(sdma_boundary_paused, SDHCIState),
         VMSTATE_END_OF_LIST()
     },
     .subsections = (const VMStateDescription * const []) {
include/hw/sd/sdhci.h +2 -0
index a9da6203fc..c485277744 100644
@@ -103,6 +103,8 @@ struct SDHCIState {
      * to be protected. Set wp_inverted to invert the signal.
      */
     bool wp_inverted;
+    /* Indicate that SDMA transfer is paused due to hitting the boundary */
+    bool sdma_boundary_paused;
 };
 typedef struct SDHCIState SDHCIState;
 
Signature · 1 line
2.43.0
#2 lore raw

Re: [PATCH v3 1/2] hw/sd: sdhci: Fix SDMA boundary bug

Quoted 110 lines · Tao Ding

During SDMA transfers, the controller raises an interrupt at buffer boundaries to request a system address update. The host driver will rewrite the SDHC_SYSAD register. (according to PartA2_SD_Host_Controller_Simplified_Specification_Ver2.00.pdf section 2.2.1) However, the current code will ignore write operations to SDHC_SYSAD.

To fix this bug, when SDMA encounters a boundary, a state is set to record it.
In this state, SDHC_SYSAD can be written. In other cases, it is still protected by TRANSFERRING_DATA.

---
The reviewer pointed out an issue in patch v2 and suggested a more appropriate way of modification. In this patch, a new state has been added to indicate encountering a boundary.
No additional state has been added to the existing stopped_state,
which seems to be specifically for gap requests (as can be seen from some if-else statements in the code)

This bug may be encountered when using uboot with SDMA functionality, and can be reproduced by following the steps below. Uboot test:
1. Prepare zImage, uboot and rootfs.cpio.gz, make sure uboot config (MMC_SDHCI_SDMA=y).
for more detail, can obtained "https://lore.kernel.org/qemu-devel/[email protected]/T/#mcdf231e6bd762903f8393bf93253cbcb12ececcf"
2. ./qemu-system-aarch64 -M xilinx-zynq-a9 -machine boot-mode=sd -m 1024 -display none -serial null -serial stdio -monitor none \
-device loader,file=u-boot-dtb.bin,addr=0x04000000,cpu-num=0 \
-drive file=zynq-sd.img,format=raw,if=sd -D qemu.log

After applying the fix, success load sd card:

U-Boot 2026.07-rc4-g1e80ee41441c (Jun 15 2026 - 19:05:05 +0800)

Model: Xilinx ZC702 board
DRAM:  ECC disabled 1 GiB
Core:  33 devices, 21 uclasses, devicetree: board
Flash: 0 Bytes
NAND:  0 MiB
MMC:   mmc@e0100000: 0

Loading Environment from FAT... *** Error - No Valid Environment Area found *** Warning - bad env area, using default environment

In:    serial@e0001000
Out:   serial@e0001000
Err:   serial@e0001000

Net:
ZYNQ GEM: e000b000, mdio bus e000b000, phyaddr 7, interface rgmii-id

Warning: ethernet@e000b000 (eth0) using random MAC address - 4a:63:5a:a4:d3:6b
eth0: ethernet@e000b000
Hit any key to stop autoboot: 0
Zynq> fatload mmc 0 ${kernel_addr_r} zImage
10838528 bytes read in 4316 ms (2.4 MiB/s)
Zynq> fatload mmc 0 ${ramdisk_addr_r} rootfs-arm32.cpio.gz
406531 bytes read in 199 ms (1.9 MiB/s)
Zynq> setenv ramdisk_size ${filesize}
Zynq> fatload mmc 0 ${fdt_addr_r} zynq-zc702.dtb
15754 bytes read in 21 ms (732.4 KiB/s)
Zynq> setenv bootargs console=ttyPS0,115200 earlycon ignore_loglevel rdinit=/init Zynq> bootz ${kernel_addr_r} ${ramdisk_addr_r}:${ramdisk_size} ${fdt_addr_r} Kernel image @ 0x2000000 [ 0x000000 - 0xa56200 ]
## Flattened Device Tree blob at 01f00000
Booting using the fdt blob at 0x1f00000
Working FDT set to 1f00000
Loading Ramdisk to 2ff9c000, end 2ffff403 ... OK
Loading Device Tree to 2ff95000, end 2ff9bd89 ... OK
Working FDT set to 2ff95000

Starting kernel ...

 hw/sd/sdhci.c         | 9 ++++++++-
 include/hw/sd/sdhci.h | 2 ++
 2 files changed, 10 insertions(+), 1 deletion(-)
hw/sd/sdhci.c +4 -0
index c86dfa281f..ecd286f2d7 100644
@@ -307,6 +307,7 @@ static void sdhci_reset(SDHCIState *s)
     s->data_count = 0;
     s->stopped_state = sdhc_not_stopped;
     s->pending_insert_state = false;
+    s->sdma_boundary_paused = false;
     if (object_dynamic_cast(OBJECT(s), TYPE_FSL_ESDHC_BE) ||
             object_dynamic_cast(OBJECT(s), TYPE_FSL_ESDHC_LE)) {
         s->norintstsen = 0x013f;
@@ -620,6 +621,7 @@ static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s)
     }
 
     s->prnsts |= SDHC_DATA_INHIBIT | SDHC_DAT_LINE_ACTIVE;
+    s->sdma_boundary_paused = false;
     if (s->trnmod & SDHC_TRNS_READ) {
         s->prnsts |= SDHC_DOING_READ;
         while (s->blkcnt) {
@@ -644,6 +646,7 @@ static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s)
                 s->data_count = 0;
             }
             if (page_aligned && boundary_count == 0) {
+                s->sdma_boundary_paused = true;
                 break;
             }
         }
@@ -669,6 +672,7 @@ static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s)
                 }
             }
             if (page_aligned && boundary_count == 0) {
+                s->sdma_boundary_paused = true;
                 break;
             }
         }

hw/sd/sdhci.c static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s) · line 687

Quoted 5 lines
hw/sd/sdhci.c +1 -0
@@ -679,6 +683,7 @@ static void sdhci_sdma_transfer_multi_blocks(SDHCIState *s)
     }
 
     if (s->blkcnt == 0) {
+        s->sdma_boundary_paused = false;
         sdhci_end_transfer(s);

When s->blkcnt decreases to 1 the logic goes to sdhci_sdma_transfer_single_block(() where s->sdma_boundary_paused should be cleared too, otherwise it will cause potential subsequent bypass of TRANSFERRING_DATA() check in the SYSAD programming.

Quoted 20 lines
hw/sd/sdhci.c +2 -1
     } else {
         sdhci_update_irq(s);
@@ -1164,6 +1169,7 @@ static inline void sdhci_reset_write(SDHCIState *s, uint8_t value)
                 SDHC_DATA_INHIBIT | SDHC_DAT_LINE_ACTIVE);
         s->blkgap &= ~(SDHC_STOP_AT_GAP_REQ | SDHC_CONTINUE_REQ);
         s->stopped_state = sdhc_not_stopped;
+        s->sdma_boundary_paused = false;
         s->norintsts &= ~(SDHC_NIS_WBUFRDY | SDHC_NIS_RBUFRDY |
                 SDHC_NIS_DMA | SDHC_NIS_TRSCMP | SDHC_NIS_BLKGAP);
         break;
@@ -1185,7 +1191,7 @@ sdhci_write(void *opaque, hwaddr offset, uint64_t val, unsigned size)
 
     switch (offset & ~0x3) {
     case SDHC_SYSAD:
-        if (!TRANSFERRING_DATA(s->prnsts)) {
+        if (!TRANSFERRING_DATA(s->prnsts) || s->sdma_boundary_paused) {
             s->sdmasysad = (s->sdmasysad & mask) | value;
             MASKED_WRITE(s->sdmasysad, mask, value);
             /* Writing to last byte of sdmasysad might trigger transfer */

hw/sd/sdhci.c const VMStateDescription sdhci_vmstate = { · line 1517

Quoted 4 lines
hw/sd/sdhci.c +1 -0
@@ -1508,6 +1514,7 @@ const VMStateDescription sdhci_vmstate = {
         VMSTATE_VBUFFER_UINT32(fifo_buffer, SDHCIState, 1, NULL, buf_maxsz),
         VMSTATE_TIMER_PTR(insert_timer, SDHCIState),
         VMSTATE_TIMER_PTR(transfer_timer, SDHCIState),
+        VMSTATE_BOOL(sdma_boundary_paused, SDHCIState),

The version should be bumped up for migration.

Quoted 8 lines
hw/sd/sdhci.c +0 -0
         VMSTATE_END_OF_LIST()
     },
     .subsections = (const VMStateDescription * const []) {
include/hw/sd/sdhci.h +0 -0
index a9da6203fc..c485277744 100644

include/hw/sd/sdhci.h struct SDHCIState { · line 110

Quoted 8 lines
include/hw/sd/sdhci.h +2 -0
@@ -103,6 +103,8 @@ struct SDHCIState {
      * to be protected. Set wp_inverted to invert the signal.
      */
     bool wp_inverted;
+    /* Indicate that SDMA transfer is paused due to hitting the boundary */
+    bool sdma_boundary_paused;
 };
 typedef struct SDHCIState SDHCIState;
 

Regards,
Bin

author
#3 lore raw

PATCH v3 2/2 qtest: add xilinx-zynq SDHCI sdma test

In reply to Tao Ding

Add a standalone qtest for the SDHCI controller on the xilinx-zynq-a9 machine.

The test covers the SDMA boundary continuation case for a multi-block read. After the controller stops at a boundary, software updates SYSAD to continue the transfer. This verifies the generic SDHCI behavior that previously failed when the continuation SYSAD write was ignored and the transfer stalled.

---
Build:
mkdir build
cd build
../configure --target-list="arm-softmmu" --disable-slirp --enable-debug --disable-werror make -j4

Run:
QTEST_QEMU_BINARY=./qemu-system-arm \
tests/qtest/xilinx-zynq-sdhci-test

Before the fix, the test fails with a BLKCNT assertion:
ERROR:../tests/qtest/xilinx-zynq-sdhci-test.c:90:wait_for_block_count:
assertion failed (qtest_readw(qts, XILINX_ZYNQ_MMC_BA + SDHC_BLKCNT) == expected): (1 == 0)

After applying the fix, the test passes:
ok 1 /arm/xilinx-zynq-sdhci/sdma/read-boundary-continue

 tests/qtest/meson.build              |   1 +
 tests/qtest/xilinx-zynq-sdhci-test.c | 178 +++++++++++++++++++++++++++
 2 files changed, 179 insertions(+)
 create mode 100644 tests/qtest/xilinx-zynq-sdhci-test.c
tests/qtest/meson.build +1 -0
index 2ae720156f..c12a35721d 100644
@@ -244,6 +244,7 @@ qtests_arm = \
    config_all_devices.has_key('CONFIG_MUSICPAL') ? ['pflash-cfi02-test'] : []) + \
   (config_all_devices.has_key('CONFIG_ASPEED_SOC') ? qtests_aspeed : []) + \
   (config_all_devices.has_key('CONFIG_NPCM7XX') ? qtests_npcm7xx : []) + \
+  (config_all_devices.has_key('CONFIG_ZYNQ') ? ['xilinx-zynq-sdhci-test'] : []) + \
   (config_all_devices.has_key('CONFIG_GENERIC_LOADER') ? ['hexloader-test'] : []) + \
   (config_all_devices.has_key('CONFIG_TPM_TIS_I2C') ? ['tpm-tis-i2c-test', 'tpm-tis-i2c-swtpm-test'] : []) + \
   (config_all_devices.has_key('CONFIG_VEXPRESS') ? ['test-arm-mptimer'] : []) + \
tests/qtest/xilinx-zynq-sdhci-test.c new file +178 -0

178 lines not on this page · raw message

Signature · 1 line
2.43.0
Reviewed-by on v3 2/2
#4 lore raw

Re: [PATCH v3 2/2] qtest: add xilinx-zynq SDHCI sdma test

Quoted 27 lines · Tao Ding

Add a standalone qtest for the SDHCI controller on the xilinx-zynq-a9 machine.

The test covers the SDMA boundary continuation case for a multi-block read. After the controller stops at a boundary, software updates SYSAD to continue the transfer. This verifies the generic SDHCI behavior that previously failed when the continuation SYSAD write was ignored and the transfer stalled.

---
Build:
mkdir build
cd build
../configure --target-list="arm-softmmu" --disable-slirp --enable-debug --disable-werror make -j4

Run:
QTEST_QEMU_BINARY=./qemu-system-arm \
tests/qtest/xilinx-zynq-sdhci-test

Before the fix, the test fails with a BLKCNT assertion:
ERROR:../tests/qtest/xilinx-zynq-sdhci-test.c:90:wait_for_block_count:
assertion failed (qtest_readw(qts, XILINX_ZYNQ_MMC_BA + SDHC_BLKCNT) == expected): (1 == 0)

Quoted 8 lines

After applying the fix, the test passes:
ok 1 /arm/xilinx-zynq-sdhci/sdma/read-boundary-continue

 tests/qtest/meson.build              |   1 +
 tests/qtest/xilinx-zynq-sdhci-test.c | 178 +++++++++++++++++++++++++++
 2 files changed, 179 insertions(+)
 create mode 100644 tests/qtest/xilinx-zynq-sdhci-test.c