Skip to contentSkip to conversation

GT64120 PCI endianness fixes and cleanup

PATCH In review Rakesh Jeyasingh 4 participants 7 messages Apr 29, 2025 – May 22, 2025

lore
Share & tools
lore
#0 lore raw

v5 cover GT64120 PCI endianness fixes and cleanup

Changes since v4:
1.Introduced needs_bswap() helper for clean endianness logic 2.use the existing pci_host_data_le_ops.read/write from hw/pci/pci_host.c

v4:https://patchew.org/QEMU/[email protected]/

Rakesh Jeyasingh (2):
hw/pci-host/gt64120: Fix endianness handling
hw/pci-host: Remove unused pci_host_data_be_ops

 hw/pci-host/gt64120.c      | 82 ++++++++++++++++++++++----------------
 hw/pci/pci_host.c          |  6 ---
 include/hw/pci-host/dino.h |  4 --
 include/hw/pci/pci_host.h  |  1 -
 4 files changed, 48 insertions(+), 45 deletions(-)
Signature · 1 line
2.43.0
#1 lore raw

PATCH v5 1/2 hw/pci-host/gt64120: Fix endianness handling

The GT-64120 PCI controller requires special handling where:
1. Host bridge(bus 0 ,device 0) must never be byte-swapped
2. Other devices follow MByteSwap bit in GT_PCI0_CMD

The previous implementation incorrectly swapped all accesses, breaking host bridge detection (lspci -d 11ab:4620).

Changes made:
1. Removed gt64120_update_pci_cfgdata_mapping() and moved data_mem initialization to gt64120_realize() for cleaner setup
2. Implemented custom read/write handlers that:
- Preserve host bridge accesses (extract32(config_reg,11,13)==0)
- apply swapping only for non-bridge devices in big-endian mode

Fixes 145e2198 ("hw/mips/gt64xxx_pci: Endian-swap using PCI_HOST_BRIDGE MemoryRegionOps")

Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2826

---
 hw/pci-host/gt64120.c | 82 +++++++++++++++++++++++++------------------
 1 file changed, 48 insertions(+), 34 deletions(-)
hw/pci-host/gt64120.c +48 -34
index 56a6ef93b7..ecb203a3d0 100644
@@ -320,38 +320,6 @@ static void gt64120_isd_mapping(GT64120State *s)
     memory_region_transaction_commit();
 }
 
-static void gt64120_update_pci_cfgdata_mapping(GT64120State *s)
-{
-    /* Indexed on MByteSwap bit, see Table 158: PCI_0 Command, Offset: 0xc00 */
-    static const MemoryRegionOps *pci_host_data_ops[] = {
-        &pci_host_data_be_ops, &pci_host_data_le_ops
-    };
-    PCIHostState *phb = PCI_HOST_BRIDGE(s);
-
-    memory_region_transaction_begin();
-
-    /*
-     * The setting of the MByteSwap bit and MWordSwap bit in the PCI Internal
-     * Command Register determines how data transactions from the CPU to/from
-     * PCI are handled along with the setting of the Endianness bit in the CPU
-     * Configuration Register. See:
-     * - Table 16: 32-bit PCI Transaction Endianness
-     * - Table 158: PCI_0 Command, Offset: 0xc00
-     */
-
-    if (memory_region_is_mapped(&phb->data_mem)) {
-        memory_region_del_subregion(&s->ISD_mem, &phb->data_mem);
-        object_unparent(OBJECT(&phb->data_mem));
-    }
-    memory_region_init_io(&phb->data_mem, OBJECT(phb),
-                          pci_host_data_ops[s->regs[GT_PCI0_CMD] & 1],
-                          s, "pci-conf-data", 4);
-    memory_region_add_subregion_overlap(&s->ISD_mem, GT_PCI0_CFGDATA << 2,
-                                        &phb->data_mem, 1);
-
-    memory_region_transaction_commit();
-}
-
 static void gt64120_pci_mapping(GT64120State *s)
 {
     memory_region_transaction_begin();
@@ -645,7 +613,6 @@ static void gt64120_writel(void *opaque, hwaddr addr,
     case GT_PCI0_CMD:
     case GT_PCI1_CMD:
         s->regs[saddr] = val & 0x0401fc0f;
-        gt64120_update_pci_cfgdata_mapping(s);
         break;
     case GT_PCI0_TOR:
     case GT_PCI0_BS_SCS10:
@@ -1024,6 +991,48 @@ static const MemoryRegionOps isd_mem_ops = {
     },
 };
 
+static bool bswap(const GT64120State *s) 
+{
+    PCIHostState *phb = PCI_HOST_BRIDGE(s);
+    /*check for bus == 0 && device == 0, Bits 11:15 = Device , Bits 16:23 = Bus*/
+    bool is_phb_dev0 = extract32(phb->config_reg, 11, 13) == 0;
+    bool le_mode = FIELD_EX32(s->regs[GT_PCI0_CMD], GT_PCI0_CMD, MByteSwap);
+    /* Only swap for non-bridge devices in big-endian mode */
+    return !le_mode && !is_phb_dev0;
+}
+
+static uint64_t gt64120_pci_data_read(void *opaque, hwaddr addr, unsigned size)
+{
+    GT64120State *s = opaque;
+    uint64_t val = pci_host_data_le_ops.read(opaque, addr, size);
+
+    if (bswap(s)) {
+        val = bswap32(val);
+    }
+    return val;
+}
+
+static void gt64120_pci_data_write(void *opaque, hwaddr addr, 
+    uint64_t val, unsigned size)
+{
+    GT64120State *s = opaque;
+
+    if (bswap(s)) {
+        val = bswap32(val); 
+    }
+    pci_host_data_le_ops.write(opaque, addr, val, size);  
+}
+
+static const MemoryRegionOps gt64120_pci_data_ops = {
+    .read = gt64120_pci_data_read,
+    .write = gt64120_pci_data_write,
+    .endianness = DEVICE_LITTLE_ENDIAN,
+    .valid = {
+        .min_access_size = 4,
+        .max_access_size = 4,
+    },
+};
+
 static void gt64120_reset(DeviceState *dev)
 {
     GT64120State *s = GT64120_PCI_HOST_BRIDGE(dev);
@@ -1178,7 +1187,6 @@ static void gt64120_reset(DeviceState *dev)
 
     gt64120_isd_mapping(s);
     gt64120_pci_mapping(s);
-    gt64120_update_pci_cfgdata_mapping(s);
 }
 
 static void gt64120_realize(DeviceState *dev, Error **errp)
@@ -1202,6 +1210,12 @@ static void gt64120_realize(DeviceState *dev, Error **errp)
     memory_region_add_subregion_overlap(&s->ISD_mem, GT_PCI0_CFGADDR << 2,
                                         &phb->conf_mem, 1);
 
+    memory_region_init_io(&phb->data_mem, OBJECT(phb),
+                          &gt64120_pci_data_ops,
+                          s, "pci-conf-data", 4);
+    memory_region_add_subregion_overlap(&s->ISD_mem, GT_PCI0_CFGDATA << 2,
+                                        &phb->data_mem, 1);
+
 
     /*
      * The whole address space decoded by the GT-64120A doesn't generate
Signature · 1 line
2.43.0
#2 lore raw

Re: [PATCH v5 1/2] hw/pci-host/gt64120: Fix endianness handling

Quoted 98 lines · Rakesh Jeyasingh

The GT-64120 PCI controller requires special handling where:
1. Host bridge(bus 0 ,device 0) must never be byte-swapped
2. Other devices follow MByteSwap bit in GT_PCI0_CMD

The previous implementation incorrectly swapped all accesses, breaking host bridge detection (lspci -d 11ab:4620).

Changes made:
1. Removed gt64120_update_pci_cfgdata_mapping() and moved data_mem initialization
to gt64120_realize() for cleaner setup
2. Implemented custom read/write handlers that:
- Preserve host bridge accesses (extract32(config_reg,11,13)==0)
- apply swapping only for non-bridge devices in big-endian mode

Fixes 145e2198 ("hw/mips/gt64xxx_pci: Endian-swap using PCI_HOST_BRIDGE

MemoryRegionOps")
Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2826

---
 hw/pci-host/gt64120.c | 82 +++++++++++++++++++++++++------------------
 1 file changed, 48 insertions(+), 34 deletions(-)
hw/pci-host/gt64120.c +0 -3
index 56a6ef93b7..ecb203a3d0 100644
@@ -320,38 +320,6 @@ static void gt64120_isd_mapping(GT64120State *s)
     memory_region_transaction_commit();
 }
 
-static void gt64120_update_pci_cfgdata_mapping(GT64120State *s)
-{
-    /* Indexed on MByteSwap bit, see Table 158: PCI_0 Command, Offset:

0xc00 */

hw/pci-host/gt64120.c +0 -9
-    static const MemoryRegionOps *pci_host_data_ops[] = {
-        &pci_host_data_be_ops, &pci_host_data_le_ops
-    };
-    PCIHostState *phb = PCI_HOST_BRIDGE(s);
-
-    memory_region_transaction_begin();
-
-    /*
-     * The setting of the MByteSwap bit and MWordSwap bit in the PCI

Internal
- * Command Register determines how data transactions from the CPU
to/from
- * PCI are handled along with the setting of the Endianness bit in
the CPU

hw/pci-host/gt64120.c +4 -19
-     * Configuration Register. See:
-     * - Table 16: 32-bit PCI Transaction Endianness
-     * - Table 158: PCI_0 Command, Offset: 0xc00
-     */
-
-    if (memory_region_is_mapped(&phb->data_mem)) {
-        memory_region_del_subregion(&s->ISD_mem, &phb->data_mem);
-        object_unparent(OBJECT(&phb->data_mem));
-    }
-    memory_region_init_io(&phb->data_mem, OBJECT(phb),
-                          pci_host_data_ops[s->regs[GT_PCI0_CMD] & 1],
-                          s, "pci-conf-data", 4);
-    memory_region_add_subregion_overlap(&s->ISD_mem, GT_PCI0_CFGDATA << 2,
-                                        &phb->data_mem, 1);
-
-    memory_region_transaction_commit();
-}
-
 static void gt64120_pci_mapping(GT64120State *s)
 {
     memory_region_transaction_begin();
@@ -645,7 +613,6 @@ static void gt64120_writel(void *opaque, hwaddr addr,
     case GT_PCI0_CMD:
     case GT_PCI1_CMD:
         s->regs[saddr] = val & 0x0401fc0f;
-        gt64120_update_pci_cfgdata_mapping(s);
         break;
     case GT_PCI0_TOR:
     case GT_PCI0_BS_SCS10:
@@ -1024,6 +991,48 @@ static const MemoryRegionOps isd_mem_ops = {
     },
 };
 
+static bool bswap(const GT64120State *s)
+{
+    PCIHostState *phb = PCI_HOST_BRIDGE(s);
+    /*check for bus == 0 && device == 0, Bits 11:15 = Device , Bits 16:23

= Bus*/

hw/pci-host/gt64120.c +2 -0
+    bool is_phb_dev0 = extract32(phb->config_reg, 11, 13) == 0;
+    bool le_mode = FIELD_EX32(s->regs[GT_PCI0_CMD], GT_PCI0_CMD,

MByteSwap);

hw/pci-host/gt64120.c +5 -0
+    /* Only swap for non-bridge devices in big-endian mode */
+    return !le_mode && !is_phb_dev0;
+}
+
+static uint64_t gt64120_pci_data_read(void *opaque, hwaddr addr, unsigned

size)

hw/pci-host/gt64120.c static const MemoryRegionOps isd_mem_ops = { · line 1007

Quoted 3 lines
hw/pci-host/gt64120.c +3 -0
+{
+    GT64120State *s = opaque;
+    uint64_t val = pci_host_data_le_ops.read(opaque, addr, size);

Hi , I just noticed that I made a mistake in the declaring data length of read val. In previous commits declared as uint32_t (as I think the PCI controller handles 32-bit values, while MemoryRegionOps uses uint64_t returns)
Should i consider:
1.sending corrected patch(hoping it not yet merged)
2.or any suggestion on fixing it?

Quoted 57 lines
hw/pci-host/gt64120.c +34 -1
+
+    if (bswap(s)) {
+        val = bswap32(val);
+    }
+    return val;
+}
+
+static void gt64120_pci_data_write(void *opaque, hwaddr addr,
+    uint64_t val, unsigned size)
+{
+    GT64120State *s = opaque;
+
+    if (bswap(s)) {
+        val = bswap32(val);
+    }
+    pci_host_data_le_ops.write(opaque, addr, val, size);
+}
+
+static const MemoryRegionOps gt64120_pci_data_ops = {
+    .read = gt64120_pci_data_read,
+    .write = gt64120_pci_data_write,
+    .endianness = DEVICE_LITTLE_ENDIAN,
+    .valid = {
+        .min_access_size = 4,
+        .max_access_size = 4,
+    },
+};
+
 static void gt64120_reset(DeviceState *dev)
 {
     GT64120State *s = GT64120_PCI_HOST_BRIDGE(dev);
@@ -1178,7 +1187,6 @@ static void gt64120_reset(DeviceState *dev)
 
     gt64120_isd_mapping(s);
     gt64120_pci_mapping(s);
-    gt64120_update_pci_cfgdata_mapping(s);
 }
 
 static void gt64120_realize(DeviceState *dev, Error **errp)
@@ -1202,6 +1210,12 @@ static void gt64120_realize(DeviceState *dev, Error
 **errp)
     memory_region_add_subregion_overlap(&s->ISD_mem, GT_PCI0_CFGADDR << 2,
                                         &phb->conf_mem, 1);
 
+    memory_region_init_io(&phb->data_mem, OBJECT(phb),
+                          &gt64120_pci_data_ops,
+                          s, "pci-conf-data", 4);
+    memory_region_add_subregion_overlap(&s->ISD_mem, GT_PCI0_CFGDATA << 2,
+                                        &phb->data_mem, 1);
+
 
     /*

* The whole address space decoded by the GT-64120A doesn't generate
--
2.43.0

#3 lore raw

Re: [PATCH v5 1/2] hw/pci-host/gt64120: Fix endianness handling

Quoted 109 lines · Rakesh Jeyasingh

On Tue, Apr 29, 2025 at 10:34 PM Rakesh Jeyasingh <[email protected] <mailto:[email protected]>> wrote:

    The GT-64120 PCI controller requires special handling where:
    1. Host bridge(bus 0 ,device 0) must never be byte-swapped
    2. Other devices follow MByteSwap bit in GT_PCI0_CMD

    The previous implementation incorrectly  swapped all accesses, breaking
    host bridge detection (lspci -d 11ab:4620).

    Changes made:
    1. Removed gt64120_update_pci_cfgdata_mapping() and moved data_mem
    initialization
       to gt64120_realize() for cleaner setup
    2. Implemented custom read/write handlers that:
        - Preserve host bridge accesses (extract32(config_reg,11,13)==0)
        - apply swapping only for non-bridge devices in big-endian mode

    Fixes: 145e2198 ("hw/mips/gt64xxx_pci: Endian-swap using
    PCI_HOST_BRIDGE MemoryRegionOps")
    Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2826
    <https://gitlab.com/qemu-project/qemu/-/issues/2826>

    Signed-off-by: Rakesh Jeyasingh <[email protected]
    <mailto:[email protected]>>
    ---
      hw/pci-host/gt64120.c | 82 +++++++++++++++++++++++++------------------
      1 file changed, 48 insertions(+), 34 deletions(-)

    diff --git a/hw/pci-host/gt64120.c b/hw/pci-host/gt64120.c
    index 56a6ef93b7..ecb203a3d0 100644
    --- a/hw/pci-host/gt64120.c
    +++ b/hw/pci-host/gt64120.c
    @@ -320,38 +320,6 @@ static void gt64120_isd_mapping(GT64120State *s)
          memory_region_transaction_commit();
      }

    -static void gt64120_update_pci_cfgdata_mapping(GT64120State *s)
    -{
    -    /* Indexed on MByteSwap bit, see Table 158: PCI_0 Command,
    Offset: 0xc00 */
    -    static const MemoryRegionOps *pci_host_data_ops[] = {
    -        &pci_host_data_be_ops, &pci_host_data_le_ops
    -    };
    -    PCIHostState *phb = PCI_HOST_BRIDGE(s);
    -
    -    memory_region_transaction_begin();
    -
    -    /*
    -     * The setting of the MByteSwap bit and MWordSwap bit in the
    PCI Internal
    -     * Command Register determines how data transactions from the
    CPU to/from
    -     * PCI are handled along with the setting of the Endianness bit
    in the CPU
    -     * Configuration Register. See:
    -     * - Table 16: 32-bit PCI Transaction Endianness
    -     * - Table 158: PCI_0 Command, Offset: 0xc00
    -     */
    -
    -    if (memory_region_is_mapped(&phb->data_mem)) {
    -        memory_region_del_subregion(&s->ISD_mem, &phb->data_mem);
    -        object_unparent(OBJECT(&phb->data_mem));
    -    }
    -    memory_region_init_io(&phb->data_mem, OBJECT(phb),
    -                          pci_host_data_ops[s->regs[GT_PCI0_CMD] & 1],
    -                          s, "pci-conf-data", 4);
    -    memory_region_add_subregion_overlap(&s->ISD_mem,
    GT_PCI0_CFGDATA << 2,
    -                                        &phb->data_mem, 1);
    -
    -    memory_region_transaction_commit();
    -}
    -
      static void gt64120_pci_mapping(GT64120State *s)
      {
          memory_region_transaction_begin();
    @@ -645,7 +613,6 @@ static void gt64120_writel(void *opaque, hwaddr
    addr,
          case GT_PCI0_CMD:
          case GT_PCI1_CMD:
              s->regs[saddr] = val & 0x0401fc0f;
    -        gt64120_update_pci_cfgdata_mapping(s);
              break;
          case GT_PCI0_TOR:
          case GT_PCI0_BS_SCS10:
    @@ -1024,6 +991,48 @@ static const MemoryRegionOps isd_mem_ops = {
          },
      };

    +static bool bswap(const GT64120State *s)
    +{
    +    PCIHostState *phb = PCI_HOST_BRIDGE(s);
    +    /*check for bus == 0 && device == 0, Bits 11:15 = Device , Bits
    16:23 = Bus*/
    +    bool is_phb_dev0 = extract32(phb->config_reg, 11, 13) == 0;
    +    bool le_mode = FIELD_EX32(s->regs[GT_PCI0_CMD], GT_PCI0_CMD,
    MByteSwap);
    +    /* Only swap for non-bridge devices in big-endian mode */
    +    return !le_mode && !is_phb_dev0;
    +}
    +
    +static uint64_t gt64120_pci_data_read(void *opaque, hwaddr addr,
    unsigned size)
    +{
    +    GT64120State *s = opaque;
    +    uint64_t val = pci_host_data_le_ops.read(opaque, addr, size);
Quoted 4 lines

Hi , I just noticed that I made a mistake in the declaring data length of read val. In previous commits declared as uint32_t (as I think the PCI controller handles 32-bit values, while MemoryRegionOps uses uint64_t returns) Should i consider:
1.sending corrected patch(hoping it not yet merged)
2.or any suggestion on fixing it?

It works with uint64_t as well, no worries. But I can also change it (I haven't applied the patch yet because I was expecting someone else to do it, but I will pick it up now).

Paolo

#4 lore raw

Re: [PATCH v5 1/2] hw/pci-host/gt64120: Fix endianness handling

In reply to Rakesh Jeyasingh
Quoted 14 lines · Rakesh Jeyasingh

The GT-64120 PCI controller requires special handling where:
1. Host bridge(bus 0 ,device 0) must never be byte-swapped
2. Other devices follow MByteSwap bit in GT_PCI0_CMD

The previous implementation incorrectly swapped all accesses, breaking host bridge detection (lspci -d 11ab:4620).

Changes made:
1. Removed gt64120_update_pci_cfgdata_mapping() and moved data_mem initialization to gt64120_realize() for cleaner setup
2. Implemented custom read/write handlers that:
- Preserve host bridge accesses (extract32(config_reg,11,13)==0)
- apply swapping only for non-bridge devices in big-endian mode

Quoted 3 lines
Fixes 145e2198 ("hw/mips/gt64xxx_pci: Endian-swap using PCI_HOST_BRIDGE MemoryRegionOps")

Resolves: https://gitlab.com/qemu-project/qemu/-/issues/2826

This seems to be qemu-stable material.

I picked this one up for 9.2 and 10.0,
please let me know if I shouldn't.

Thanks,

/mjt

#5 lore raw

PATCH v5 2/2 hw/pci-host: Remove unused pci_host_data_be_ops

In reply to Rakesh Jeyasingh

pci_host_data_be_ops became unused after endianness fixes

---
 hw/pci/pci_host.c          | 6 ------
 include/hw/pci-host/dino.h | 4 ----
 include/hw/pci/pci_host.h  | 1 -
 3 files changed, 11 deletions(-)
hw/pci/pci_host.c +0 -6
index abe83bbab8..7179d99178 100644
@@ -217,12 +217,6 @@ const MemoryRegionOps pci_host_data_le_ops = {
     .endianness = DEVICE_LITTLE_ENDIAN,
 };
 
-const MemoryRegionOps pci_host_data_be_ops = {
-    .read = pci_host_data_read,
-    .write = pci_host_data_write,
-    .endianness = DEVICE_BIG_ENDIAN,
-};
-
 static bool pci_host_needed(void *opaque)
 {
     PCIHostState *s = opaque;
include/hw/pci-host/dino.h +0 -4
index fd7975c798..5dc8cdf610 100644
@@ -109,10 +109,6 @@ static const uint32_t reg800_keep_bits[DINO800_REGS] = {
 struct DinoState {
     PCIHostState parent_obj;
 
-    /*
-     * PCI_CONFIG_ADDR is parent_obj.config_reg, via pci_host_conf_be_ops,
-     * so that we can map PCI_CONFIG_DATA to pci_host_data_be_ops.
-     */
     uint32_t config_reg_dino; /* keep original copy, including 2 lowest bits */
 
     uint32_t iar0;
include/hw/pci/pci_host.h +0 -1
index e52d8ec2cd..954dd446fa 100644
@@ -68,6 +68,5 @@ uint32_t pci_data_read(PCIBus *s, uint32_t addr, unsigned len);
 extern const MemoryRegionOps pci_host_conf_le_ops;
 extern const MemoryRegionOps pci_host_conf_be_ops;
 extern const MemoryRegionOps pci_host_data_le_ops;
-extern const MemoryRegionOps pci_host_data_be_ops;
 
 #endif /* PCI_HOST_H */
Signature · 1 line
2.43.0
Tested-by on the v5 series
#6 lore raw
In reply to Rakesh Jeyasingh
Quoted 10 lines · Rakesh Jeyasingh

Changes since v4:
1.Introduced needs_bswap() helper for clean endianness logic 2.use the existing pci_host_data_le_ops.read/write from hw/pci/pci_host.c

v4:https://patchew.org/QEMU/[email protected]/

Rakesh Jeyasingh (2):
hw/pci-host/gt64120: Fix endianness handling hw/pci-host: Remove unused pci_host_data_be_ops

Quoted 5 lines
  hw/pci-host/gt64120.c      | 82 ++++++++++++++++++++++----------------
  hw/pci/pci_host.c          |  6 ---
  include/hw/pci-host/dino.h |  4 --
  include/hw/pci/pci_host.h  |  1 -
  4 files changed, 48 insertions(+), 45 deletions(-)

Thanks for tackling this ... I gave it a try on both, little and big endian host and it seems to work as expected now.

Series