Skip to contentSkip to conversation

Fix/add vmstate handling in some I2C code

PATCH Superseded [email protected] 5 participants 19 messages Nov 7, 2018 – Nov 12, 2018

v1 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

cover Fix/add vmstate handling in some I2C code

These changes allow SMBus access while doing a state transfer. Seems like a good idea to me in general.

I have these queued for the SMBus IPMI driver work, of course.

#1 lore raw

PATCH 1/3 i2c:pm_smbus: Fix state transfer

From: Corey Minyard <[email protected]>

Transfer the state information for the SMBus registers and internal data so it will work on a VM transfer.

Cc Michael S. Tsirkin <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
 hw/acpi/piix4.c           |  3 ++-
 hw/i2c/pm_smbus.c         | 20 ++++++++++++++++++++
 hw/i2c/smbus_ich9.c       |  5 +++--
 include/hw/i2c/pm_smbus.h |  2 ++
 4 files changed, 27 insertions(+), 3 deletions(-)
hw/acpi/piix4.c +2 -1
index e330f24c71..313305f5a0 100644
@@ -309,7 +309,7 @@ static const VMStateDescription vmstate_cpuhp_state = {
  */
 static const VMStateDescription vmstate_acpi = {
     .name = "piix4_pm",
-    .version_id = 3,
+    .version_id = 4,
     .minimum_version_id = 3,
     .minimum_version_id_old = 1,
     .load_state_old = acpi_load_old,
@@ -320,6 +320,7 @@ static const VMStateDescription vmstate_acpi = {
         VMSTATE_UINT16(ar.pm1.evt.en, PIIX4PMState),
         VMSTATE_UINT16(ar.pm1.cnt.cnt, PIIX4PMState),
         VMSTATE_STRUCT(apm, PIIX4PMState, 0, vmstate_apm, APMState),
+        VMSTATE_STRUCT(smb, PIIX4PMState, 4, pmsmb_vmstate, PMSMBus),
         VMSTATE_TIMER_PTR(ar.tmr.timer, PIIX4PMState),
         VMSTATE_INT64(ar.tmr.overflow_time, PIIX4PMState),
         VMSTATE_STRUCT(ar.gpe, PIIX4PMState, 2, vmstate_gpe, ACPIGPE),
hw/i2c/pm_smbus.c +20 -0
index 685a2378ed..4d021bb017 100644
@@ -402,6 +402,26 @@ static const MemoryRegionOps pm_smbus_ops = {
     .endianness = DEVICE_LITTLE_ENDIAN,
 };
 
+const VMStateDescription pmsmb_vmstate = {
+    .name = "pmsmb",
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields = (VMStateField[]) {
+        VMSTATE_UINT8(smb_stat, PMSMBus),
+        VMSTATE_UINT8(smb_ctl, PMSMBus),
+        VMSTATE_UINT8(smb_cmd, PMSMBus),
+        VMSTATE_UINT8(smb_addr, PMSMBus),
+        VMSTATE_UINT8(smb_data0, PMSMBus),
+        VMSTATE_UINT8(smb_data1, PMSMBus),
+        VMSTATE_UINT32(smb_index, PMSMBus),
+        VMSTATE_UINT8_ARRAY(smb_data, PMSMBus, PM_SMBUS_MAX_MSG_SIZE),
+        VMSTATE_UINT8(smb_auxctl, PMSMBus),
+        VMSTATE_BOOL(i2c_enable, PMSMBus),
+        VMSTATE_BOOL(op_done, PMSMBus),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
 void pm_smbus_init(DeviceState *parent, PMSMBus *smb, bool force_aux_blk)
 {
     smb->op_done = true;
hw/i2c/smbus_ich9.c +3 -2
index 2a8b49e02f..4202b71da5 100644
@@ -47,10 +47,11 @@ typedef struct ICH9SMBState {
 
 static const VMStateDescription vmstate_ich9_smbus = {
     .name = "ich9_smb",
-    .version_id = 1,
+    .version_id = 2,
     .minimum_version_id = 1,
     .fields = (VMStateField[]) {
-        VMSTATE_PCI_DEVICE(dev, struct ICH9SMBState),
+        VMSTATE_PCI_DEVICE(dev, ICH9SMBState),
+        VMSTATE_STRUCT(smb, ICH9SMBState, 2, pmsmb_vmstate, PMSMBus),
         VMSTATE_END_OF_LIST()
     }
 };
include/hw/i2c/pm_smbus.h +2 -0
index 060d3c6ac0..471345eaa2 100644
@@ -35,4 +35,6 @@ typedef struct PMSMBus {
 
 void pm_smbus_init(DeviceState *parent, PMSMBus *smb, bool force_aux_blk);
 
+extern const VMStateDescription pmsmb_vmstate;
+
 #endif /* PM_SMBUS_H */
Signature · 1 line
2.17.1
#2 lore raw

Re: [Qemu-devel] [PATCH 1/3] i2c:pm_smbus: Fix state transfer

Quoted 4 lines · * [email protected] ([email protected]) wrote:

From: Corey Minyard <[email protected]>

Transfer the state information for the SMBus registers and internal data so it will work on a VM transfer.

Weren't the comments from August were that we'd prefer to see this in a subsection so we didn't break migration compatibility.

Dave

Quoted 95 lines
Cc Michael S. Tsirkin <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
 hw/acpi/piix4.c           |  3 ++-
 hw/i2c/pm_smbus.c         | 20 ++++++++++++++++++++
 hw/i2c/smbus_ich9.c       |  5 +++--
 include/hw/i2c/pm_smbus.h |  2 ++
 4 files changed, 27 insertions(+), 3 deletions(-)
hw/acpi/piix4.c +2 -1
index e330f24c71..313305f5a0 100644
@@ -309,7 +309,7 @@ static const VMStateDescription vmstate_cpuhp_state = {
  */
 static const VMStateDescription vmstate_acpi = {
     .name = "piix4_pm",
-    .version_id = 3,
+    .version_id = 4,
     .minimum_version_id = 3,
     .minimum_version_id_old = 1,
     .load_state_old = acpi_load_old,
@@ -320,6 +320,7 @@ static const VMStateDescription vmstate_acpi = {
         VMSTATE_UINT16(ar.pm1.evt.en, PIIX4PMState),
         VMSTATE_UINT16(ar.pm1.cnt.cnt, PIIX4PMState),
         VMSTATE_STRUCT(apm, PIIX4PMState, 0, vmstate_apm, APMState),
+        VMSTATE_STRUCT(smb, PIIX4PMState, 4, pmsmb_vmstate, PMSMBus),
         VMSTATE_TIMER_PTR(ar.tmr.timer, PIIX4PMState),
         VMSTATE_INT64(ar.tmr.overflow_time, PIIX4PMState),
         VMSTATE_STRUCT(ar.gpe, PIIX4PMState, 2, vmstate_gpe, ACPIGPE),
hw/i2c/pm_smbus.c +20 -0
index 685a2378ed..4d021bb017 100644
@@ -402,6 +402,26 @@ static const MemoryRegionOps pm_smbus_ops = {
     .endianness = DEVICE_LITTLE_ENDIAN,
 };
 
+const VMStateDescription pmsmb_vmstate = {
+    .name = "pmsmb",
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields = (VMStateField[]) {
+        VMSTATE_UINT8(smb_stat, PMSMBus),
+        VMSTATE_UINT8(smb_ctl, PMSMBus),
+        VMSTATE_UINT8(smb_cmd, PMSMBus),
+        VMSTATE_UINT8(smb_addr, PMSMBus),
+        VMSTATE_UINT8(smb_data0, PMSMBus),
+        VMSTATE_UINT8(smb_data1, PMSMBus),
+        VMSTATE_UINT32(smb_index, PMSMBus),
+        VMSTATE_UINT8_ARRAY(smb_data, PMSMBus, PM_SMBUS_MAX_MSG_SIZE),
+        VMSTATE_UINT8(smb_auxctl, PMSMBus),
+        VMSTATE_BOOL(i2c_enable, PMSMBus),
+        VMSTATE_BOOL(op_done, PMSMBus),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
 void pm_smbus_init(DeviceState *parent, PMSMBus *smb, bool force_aux_blk)
 {
     smb->op_done = true;
hw/i2c/smbus_ich9.c +3 -2
index 2a8b49e02f..4202b71da5 100644
@@ -47,10 +47,11 @@ typedef struct ICH9SMBState {
 
 static const VMStateDescription vmstate_ich9_smbus = {
     .name = "ich9_smb",
-    .version_id = 1,
+    .version_id = 2,
     .minimum_version_id = 1,
     .fields = (VMStateField[]) {
-        VMSTATE_PCI_DEVICE(dev, struct ICH9SMBState),
+        VMSTATE_PCI_DEVICE(dev, ICH9SMBState),
+        VMSTATE_STRUCT(smb, ICH9SMBState, 2, pmsmb_vmstate, PMSMBus),
         VMSTATE_END_OF_LIST()
     }
 };
include/hw/i2c/pm_smbus.h +2 -0
index 060d3c6ac0..471345eaa2 100644
@@ -35,4 +35,6 @@ typedef struct PMSMBus {
 
 void pm_smbus_init(DeviceState *parent, PMSMBus *smb, bool force_aux_blk);
 
+extern const VMStateDescription pmsmb_vmstate;
+
 #endif /* PM_SMBUS_H */
Signature · 1 line
2.17.1

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

#3 lore raw

Re: [Qemu-devel] [PATCH 1/3] i2c:pm_smbus: Fix state transfer

Quoted 7 lines · Dr. David Alan Gilbert
Quoted 4 lines · * [email protected] ([email protected]) wrote:

From: Corey Minyard <[email protected]>

Transfer the state information for the SMBus registers and internal data so it will work on a VM transfer.

Weren't the comments from August were that we'd prefer to see this in a subsection so we didn't break migration compatibility.

Yeah, I screwed up here, but I realized what happened.  I started working this some more and discovered that the eeprom didn't work on q35 at all, and while working on that I found some other issues, and then I had another critical thing come up and this got lost.  I apologize, I should have looked back.

I will go back and address your comments from the previous emails once I get everything sorted out.

Thanks,

-corey

Quoted 99 lines

Dave

Quoted 95 lines
Cc Michael S. Tsirkin <[email protected]>
Cc Paolo Bonzini <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
  hw/acpi/piix4.c           |  3 ++-
  hw/i2c/pm_smbus.c         | 20 ++++++++++++++++++++
  hw/i2c/smbus_ich9.c       |  5 +++--
  include/hw/i2c/pm_smbus.h |  2 ++
  4 files changed, 27 insertions(+), 3 deletions(-)
hw/acpi/piix4.c +2 -1
index e330f24c71..313305f5a0 100644
@@ -309,7 +309,7 @@ static const VMStateDescription vmstate_cpuhp_state = {
   */
  static const VMStateDescription vmstate_acpi = {
      .name = "piix4_pm",
-    .version_id = 3,
+    .version_id = 4,
      .minimum_version_id = 3,
      .minimum_version_id_old = 1,
      .load_state_old = acpi_load_old,
@@ -320,6 +320,7 @@ static const VMStateDescription vmstate_acpi = {
          VMSTATE_UINT16(ar.pm1.evt.en, PIIX4PMState),
          VMSTATE_UINT16(ar.pm1.cnt.cnt, PIIX4PMState),
          VMSTATE_STRUCT(apm, PIIX4PMState, 0, vmstate_apm, APMState),
+        VMSTATE_STRUCT(smb, PIIX4PMState, 4, pmsmb_vmstate, PMSMBus),
          VMSTATE_TIMER_PTR(ar.tmr.timer, PIIX4PMState),
          VMSTATE_INT64(ar.tmr.overflow_time, PIIX4PMState),
          VMSTATE_STRUCT(ar.gpe, PIIX4PMState, 2, vmstate_gpe, ACPIGPE),
hw/i2c/pm_smbus.c +20 -0
index 685a2378ed..4d021bb017 100644
@@ -402,6 +402,26 @@ static const MemoryRegionOps pm_smbus_ops = {
      .endianness = DEVICE_LITTLE_ENDIAN,
  };
 
+const VMStateDescription pmsmb_vmstate = {
+    .name = "pmsmb",
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields = (VMStateField[]) {
+        VMSTATE_UINT8(smb_stat, PMSMBus),
+        VMSTATE_UINT8(smb_ctl, PMSMBus),
+        VMSTATE_UINT8(smb_cmd, PMSMBus),
+        VMSTATE_UINT8(smb_addr, PMSMBus),
+        VMSTATE_UINT8(smb_data0, PMSMBus),
+        VMSTATE_UINT8(smb_data1, PMSMBus),
+        VMSTATE_UINT32(smb_index, PMSMBus),
+        VMSTATE_UINT8_ARRAY(smb_data, PMSMBus, PM_SMBUS_MAX_MSG_SIZE),
+        VMSTATE_UINT8(smb_auxctl, PMSMBus),
+        VMSTATE_BOOL(i2c_enable, PMSMBus),
+        VMSTATE_BOOL(op_done, PMSMBus),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
  void pm_smbus_init(DeviceState *parent, PMSMBus *smb, bool force_aux_blk)
  {
      smb->op_done = true;
hw/i2c/smbus_ich9.c +3 -2
index 2a8b49e02f..4202b71da5 100644
@@ -47,10 +47,11 @@ typedef struct ICH9SMBState {
 
  static const VMStateDescription vmstate_ich9_smbus = {
      .name = "ich9_smb",
-    .version_id = 1,
+    .version_id = 2,
      .minimum_version_id = 1,
      .fields = (VMStateField[]) {
-        VMSTATE_PCI_DEVICE(dev, struct ICH9SMBState),
+        VMSTATE_PCI_DEVICE(dev, ICH9SMBState),
+        VMSTATE_STRUCT(smb, ICH9SMBState, 2, pmsmb_vmstate, PMSMBus),
          VMSTATE_END_OF_LIST()
      }
  };
include/hw/i2c/pm_smbus.h +2 -0
index 060d3c6ac0..471345eaa2 100644
@@ -35,4 +35,6 @@ typedef struct PMSMBus {
 
  void pm_smbus_init(DeviceState *parent, PMSMBus *smb, bool force_aux_blk);
 
+extern const VMStateDescription pmsmb_vmstate;
+
  #endif /* PM_SMBUS_H */
Signature · 1 line
2.17.1

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

#4 lore raw

PATCH 2/3 i2c: Add an SMBus vmstate structure

In reply to [email protected]

From: Corey Minyard <[email protected]>

There is no vmstate handling for SMBus, so no device sitting on SMBus can have a state transfer that works reliable. So add it.

Cc Paolo Bonzini <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
 hw/i2c/smbus.c         | 14 ++++++++++++++
 include/hw/i2c/smbus.h | 18 +++++++++++++++---
 2 files changed, 29 insertions(+), 3 deletions(-)
hw/i2c/smbus.c +14 -0
index 6ff77c582f..b0774d7683 100644
@@ -349,6 +349,20 @@ int smbus_write_block(I2CBus *bus, uint8_t addr, uint8_t command, uint8_t *data,
     return 0;
 }
 
+const VMStateDescription vmstate_smbus_device = {
+    .name = TYPE_SMBUS_DEVICE,
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields      = (VMStateField[]) {
+        VMSTATE_I2C_SLAVE(i2c, SMBusDevice),
+        VMSTATE_INT32(mode, SMBusDevice),
+        VMSTATE_INT32(data_len, SMBusDevice),
+        VMSTATE_UINT8_ARRAY(data_buf, SMBusDevice, SMBUS_DATA_MAX_LEN),
+        VMSTATE_UINT8(command, SMBusDevice),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
 static void smbus_device_class_init(ObjectClass *klass, void *data)
 {
     I2CSlaveClass *sc = I2C_SLAVE_CLASS(klass);
include/hw/i2c/smbus.h +15 -3
index d8b1b9ee81..7b52020121 100644
@@ -53,14 +53,16 @@ typedef struct SMBusDeviceClass
     uint8_t (*read_data)(SMBusDevice *dev, uint8_t cmd, int n);
 } SMBusDeviceClass;
 
+#define SMBUS_DATA_MAX_LEN 34  /* command + len + 32 bytes of data.  */
+
 struct SMBusDevice {
     /* The SMBus protocol is implemented on top of I2C.  */
     I2CSlave i2c;
 
     /* Remaining fields for internal use only.  */
-    int mode;
-    int data_len;
-    uint8_t data_buf[34]; /* command + len + 32 bytes of data.  */
+    int32_t mode;
+    int32_t data_len;
+    uint8_t data_buf[SMBUS_DATA_MAX_LEN];
     uint8_t command;
 };
 
@@ -93,4 +95,14 @@ void smbus_eeprom_init_one(I2CBus *smbus, uint8_t address, uint8_t *eeprom_buf);
 void smbus_eeprom_init(I2CBus *smbus, int nb_eeprom,
                        const uint8_t *eeprom_spd, int size);
 
+extern const VMStateDescription vmstate_smbus_device;
+
+#define VMSTATE_SMBUS_DEVICE(_field, _state) {                       \
+    .name       = (stringify(_field)),                               \
+    .size       = sizeof(SMBusDevice),                               \
+    .vmsd       = &vmstate_smbus_device,                             \
+    .flags      = VMS_STRUCT,                                        \
+    .offset     = vmstate_offset_value(_state, _field, SMBusDevice), \
+}
+
 #endif
Signature · 1 line
2.17.1
#5 lore raw

Re: [Qemu-devel] [PATCH 2/3] i2c: Add an SMBus vmstate structure

Hi Corey,

Quoted 50 lines · On 7/11/18 16:54, [email protected] wrote:

From: Corey Minyard <[email protected]>

There is no vmstate handling for SMBus, so no device sitting on SMBus can have a state transfer that works reliable. So add it.

Cc Paolo Bonzini <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
  hw/i2c/smbus.c         | 14 ++++++++++++++
  include/hw/i2c/smbus.h | 18 +++++++++++++++---
  2 files changed, 29 insertions(+), 3 deletions(-)
hw/i2c/smbus.c +14 -0
index 6ff77c582f..b0774d7683 100644
@@ -349,6 +349,20 @@ int smbus_write_block(I2CBus *bus, uint8_t addr, uint8_t command, uint8_t *data,
      return 0;
  }
 
+const VMStateDescription vmstate_smbus_device = {
+    .name = TYPE_SMBUS_DEVICE,
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields      = (VMStateField[]) {
+        VMSTATE_I2C_SLAVE(i2c, SMBusDevice),
+        VMSTATE_INT32(mode, SMBusDevice),
+        VMSTATE_INT32(data_len, SMBusDevice),
+        VMSTATE_UINT8_ARRAY(data_buf, SMBusDevice, SMBUS_DATA_MAX_LEN),
+        VMSTATE_UINT8(command, SMBusDevice),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
  static void smbus_device_class_init(ObjectClass *klass, void *data)
  {
      I2CSlaveClass *sc = I2C_SLAVE_CLASS(klass);
include/hw/i2c/smbus.h +2 -0
index d8b1b9ee81..7b52020121 100644
@@ -53,14 +53,16 @@ typedef struct SMBusDeviceClass
      uint8_t (*read_data)(SMBusDevice *dev, uint8_t cmd, int n);
  } SMBusDeviceClass;
 
+#define SMBUS_DATA_MAX_LEN 34  /* command + len + 32 bytes of data.  */
+
  struct SMBusDevice {

include/hw/i2c/smbus.h typedef struct SMBusDeviceClass · line 65

Quoted 10 lines
include/hw/i2c/smbus.h +3 -3
      /* The SMBus protocol is implemented on top of I2C.  */
      I2CSlave i2c;
 
      /* Remaining fields for internal use only.  */
-    int mode;
-    int data_len;
-    uint8_t data_buf[34]; /* command + len + 32 bytes of data.  */
+    int32_t mode;
+    int32_t data_len;
+    uint8_t data_buf[SMBUS_DATA_MAX_LEN];

Those changes are not in your commit description.
Can you include them in a separate patch?
Thanks,
Phil.

Quoted 19 lines
include/hw/i2c/smbus.h +10 -0
      uint8_t command;
  };
 
@@ -93,4 +95,14 @@ void smbus_eeprom_init_one(I2CBus *smbus, uint8_t address, uint8_t *eeprom_buf);
  void smbus_eeprom_init(I2CBus *smbus, int nb_eeprom,
                         const uint8_t *eeprom_spd, int size);
 
+extern const VMStateDescription vmstate_smbus_device;
+
+#define VMSTATE_SMBUS_DEVICE(_field, _state) {                       \
+    .name       = (stringify(_field)),                               \
+    .size       = sizeof(SMBusDevice),                               \
+    .vmsd       = &vmstate_smbus_device,                             \
+    .flags      = VMS_STRUCT,                                        \
+    .offset     = vmstate_offset_value(_state, _field, SMBusDevice), \
+}
+
  #endif
#6 lore raw

Re: [Qemu-devel] [PATCH 2/3] i2c: Add an SMBus vmstate structure

Quoted 59 lines · Philippe Mathieu-Daudé

Hi Corey,

Quoted 55 lines · On 7/11/18 16:54, [email protected] wrote:

From: Corey Minyard <[email protected]>

There is no vmstate handling for SMBus, so no device sitting on SMBus can have a state transfer that works reliable. So add it.

Cc Paolo Bonzini <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
  hw/i2c/smbus.c         | 14 ++++++++++++++
  include/hw/i2c/smbus.h | 18 +++++++++++++++---
  2 files changed, 29 insertions(+), 3 deletions(-)
hw/i2c/smbus.c +13 -0
index 6ff77c582f..b0774d7683 100644
@@ -349,6 +349,20 @@ int smbus_write_block(I2CBus *bus, uint8_t addr,
 uint8_t command, uint8_t *data,
      return 0;
  }
  +const VMStateDescription vmstate_smbus_device = {
+    .name = TYPE_SMBUS_DEVICE,
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields      = (VMStateField[]) {
+        VMSTATE_I2C_SLAVE(i2c, SMBusDevice),
+        VMSTATE_INT32(mode, SMBusDevice),
+        VMSTATE_INT32(data_len, SMBusDevice),
+        VMSTATE_UINT8_ARRAY(data_buf, SMBusDevice, SMBUS_DATA_MAX_LEN),
+        VMSTATE_UINT8(command, SMBusDevice),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
  static void smbus_device_class_init(ObjectClass *klass, void *data)
  {
      I2CSlaveClass *sc = I2C_SLAVE_CLASS(klass);
include/hw/i2c/smbus.h +1 -2
index d8b1b9ee81..7b52020121 100644
@@ -53,14 +53,16 @@ typedef struct SMBusDeviceClass
      uint8_t (*read_data)(SMBusDevice *dev, uint8_t cmd, int n);
  } SMBusDeviceClass;
  +#define SMBUS_DATA_MAX_LEN 34  /* command + len + 32 bytes of data.  */
+
  struct SMBusDevice {
      /* The SMBus protocol is implemented on top of I2C.  */
      I2CSlave i2c;
        /* Remaining fields for internal use only.  */
-    int mode;
-    int data_len;
Quoted 8 lines

include/hw/i2c/smbus.h typedef struct SMBusDeviceClass · line 65

Quoted 4 lines
include/hw/i2c/smbus.h +3 -1
-    uint8_t data_buf[34]; /* command + len + 32 bytes of data.  */
+    int32_t mode;
+    int32_t data_len;
+    uint8_t data_buf[SMBUS_DATA_MAX_LEN];

Those changes are not in your commit description.
Can you include them in a separate patch?

These are just the standard "promote types in the state struct to fixed-width ones required by the VMSTATE macros", aren't they? I think they're fine as part of the vmstate conversion.

thanks
-- PMM

#7 lore raw

Re: [Qemu-devel] [PATCH 2/3] i2c: Add an SMBus vmstate structure

Quoted 69 lines · Peter Maydell
Quoted 59 lines · Philippe Mathieu-Daudé

Hi Corey,

Quoted 55 lines · On 7/11/18 16:54, [email protected] wrote:

From: Corey Minyard <[email protected]>

There is no vmstate handling for SMBus, so no device sitting on SMBus can have a state transfer that works reliable. So add it.

Cc Paolo Bonzini <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
   hw/i2c/smbus.c         | 14 ++++++++++++++
   include/hw/i2c/smbus.h | 18 +++++++++++++++---
   2 files changed, 29 insertions(+), 3 deletions(-)
hw/i2c/smbus.c +13 -0
index 6ff77c582f..b0774d7683 100644
@@ -349,6 +349,20 @@ int smbus_write_block(I2CBus *bus, uint8_t addr,
 uint8_t command, uint8_t *data,
       return 0;
   }
   +const VMStateDescription vmstate_smbus_device = {
+    .name = TYPE_SMBUS_DEVICE,
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields      = (VMStateField[]) {
+        VMSTATE_I2C_SLAVE(i2c, SMBusDevice),
+        VMSTATE_INT32(mode, SMBusDevice),
+        VMSTATE_INT32(data_len, SMBusDevice),
+        VMSTATE_UINT8_ARRAY(data_buf, SMBusDevice, SMBUS_DATA_MAX_LEN),
+        VMSTATE_UINT8(command, SMBusDevice),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
   static void smbus_device_class_init(ObjectClass *klass, void *data)
   {
       I2CSlaveClass *sc = I2C_SLAVE_CLASS(klass);
include/hw/i2c/smbus.h +1 -2
index d8b1b9ee81..7b52020121 100644
@@ -53,14 +53,16 @@ typedef struct SMBusDeviceClass
       uint8_t (*read_data)(SMBusDevice *dev, uint8_t cmd, int n);
   } SMBusDeviceClass;
   +#define SMBUS_DATA_MAX_LEN 34  /* command + len + 32 bytes of data.  */
+
   struct SMBusDevice {
       /* The SMBus protocol is implemented on top of I2C.  */
       I2CSlave i2c;
         /* Remaining fields for internal use only.  */
-    int mode;
-    int data_len;
Quoted 8 lines

include/hw/i2c/smbus.h typedef struct SMBusDeviceClass · line 65

Quoted 4 lines
include/hw/i2c/smbus.h +3 -1
-    uint8_t data_buf[34]; /* command + len + 32 bytes of data.  */
+    int32_t mode;
+    int32_t data_len;
+    uint8_t data_buf[SMBUS_DATA_MAX_LEN];

Those changes are not in your commit description.
Can you include them in a separate patch?

Quoted 3 lines

These are just the standard "promote types in the state struct to fixed-width ones required by the VMSTATE macros", aren't they? I think they're fine as part of the vmstate conversion.

Fine then!

Quoted 4 lines

thanks
-- PMM

#8 lore raw

PATCH 3/3 i2c: Add vmstate handling to the smbus eeprom

In reply to [email protected]

From: Corey Minyard <[email protected]>

This was if the eeprom is accessed during a state transfer, the transfer will be reliable.

Cc Paolo Bonzini <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
 hw/i2c/smbus_eeprom.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)
hw/i2c/smbus_eeprom.c +15 -1
index f18aa3de35..d4430b0c5d 100644
@@ -29,6 +29,8 @@
 
 //#define DEBUG
 
+#define TYPE_SMBUS_EEPROM_DEVICE "smbus-eeprom"
+
 typedef struct SMBusEEPROMDevice {
     SMBusDevice smbusdev;
     void *data;
@@ -97,6 +99,17 @@ static uint8_t eeprom_read_data(SMBusDevice *dev, uint8_t cmd, int n)
     return eeprom_receive_byte(dev);
 }
 
+static const VMStateDescription vmstate_smbus_eeprom = {
+    .name = TYPE_SMBUS_EEPROM_DEVICE,
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields      = (VMStateField[]) {
+        VMSTATE_SMBUS_DEVICE(smbusdev, SMBusEEPROMDevice),
+        VMSTATE_UINT8(offset, SMBusEEPROMDevice),
+        VMSTATE_END_OF_LIST()
+    }
+};
+
 static void smbus_eeprom_realize(DeviceState *dev, Error **errp)
 {
     SMBusEEPROMDevice *eeprom = (SMBusEEPROMDevice *)dev;
@@ -121,12 +134,13 @@ static void smbus_eeprom_class_initfn(ObjectClass *klass, void *data)
     sc->write_data = eeprom_write_data;
     sc->read_data = eeprom_read_data;
     dc->props = smbus_eeprom_properties;
+    dc->vmsd = &vmstate_smbus_eeprom;
     /* Reason: pointer property "data" */
     dc->user_creatable = false;
 }
 
 static const TypeInfo smbus_eeprom_info = {
-    .name          = "smbus-eeprom",
+    .name          = TYPE_SMBUS_EEPROM_DEVICE,
     .parent        = TYPE_SMBUS_DEVICE,
     .instance_size = sizeof(SMBusEEPROMDevice),
     .class_init    = smbus_eeprom_class_initfn,
Signature · 1 line
2.17.1
#9 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 18 lines · On 7 November 2018 at 15:54, <[email protected]> wrote:

From: Corey Minyard <[email protected]>

This was if the eeprom is accessed during a state transfer, the transfer will be reliable.

Cc Paolo Bonzini <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
 hw/i2c/smbus_eeprom.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)
hw/i2c/smbus_eeprom.c +0 -0
index f18aa3de35..d4430b0c5d 100644

hw/i2c/smbus_eeprom.c line 32

Quoted 4 lines
hw/i2c/smbus_eeprom.c +1 -0
@@ -29,6 +29,8 @@
 
 //#define DEBUG
 
+#define TYPE_SMBUS_EEPROM_DEVICE "smbus-eeprom"

The part of this patch which is adding the usual QOM macros is good, but we should provide also the cast-macro (the one that's an OBJECT_CHECK(...)). And this should be separate from adding the vmstate.

hw/i2c/smbus_eeprom.c static uint8_t eeprom_read_data(SMBusDevice *dev, uint8_t cmd, int n) · line 103

Quoted 10 lines
hw/i2c/smbus_eeprom.c +3 -0
+
 typedef struct SMBusEEPROMDevice {
     SMBusDevice smbusdev;
     void *data;
@@ -97,6 +99,17 @@ static uint8_t eeprom_read_data(SMBusDevice *dev, uint8_t cmd, int n)
     return eeprom_receive_byte(dev);
 }
 
+static const VMStateDescription vmstate_smbus_eeprom = {
+    .name = TYPE_SMBUS_EEPROM_DEVICE,

Generally we don't use the TYPE_FOO constant for the vmstate name, because we can usually freely change the TYPE_FOO string without breaking back-compat, but we can't change the vmstate name. So we decouple them to make it more obvious when a change might be changing the migration on-the-wire format.

hw/i2c/smbus_eeprom.c static uint8_t eeprom_read_data(SMBusDevice *dev, uint8_t cmd, int n) · line 111

Quoted 8 lines
hw/i2c/smbus_eeprom.c +8 -0
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields      = (VMStateField[]) {
+        VMSTATE_SMBUS_DEVICE(smbusdev, SMBusEEPROMDevice),
+        VMSTATE_UINT8(offset, SMBusEEPROMDevice),
+        VMSTATE_END_OF_LIST()
+    }
+};

This doesn't do anything for migration of the actual data contents. The current users of this device (hw/arm/aspeed.c and the smbus_eeprom_init() function) doesn't do anything
to migrate the contents. For that matter, "user of the device passes a pointer to a random lump of memory via a device property" is a bit funky as an interface. The device should allocate its own memory and migrate it, and the user should just pass the initial required contents (default being "zero-initialize", which is what everybody except the mips_fulong2e, mips_malta and sam460ex want).

Does this also break migration from an old QEMU to a new one? (For Aspeed that's probably ok, but we should flag it up in the commit message if so. x86 uses need care...)

Quoted 20 lines
hw/i2c/smbus_eeprom.c +3 -2
+
 static void smbus_eeprom_realize(DeviceState *dev, Error **errp)
 {
     SMBusEEPROMDevice *eeprom = (SMBusEEPROMDevice *)dev;
@@ -121,12 +134,13 @@ static void smbus_eeprom_class_initfn(ObjectClass *klass, void *data)
     sc->write_data = eeprom_write_data;
     sc->read_data = eeprom_read_data;
     dc->props = smbus_eeprom_properties;
+    dc->vmsd = &vmstate_smbus_eeprom;
     /* Reason: pointer property "data" */
     dc->user_creatable = false;
 }
 
 static const TypeInfo smbus_eeprom_info = {
-    .name          = "smbus-eeprom",
+    .name          = TYPE_SMBUS_EEPROM_DEVICE,
     .parent        = TYPE_SMBUS_DEVICE,
     .instance_size = sizeof(SMBusEEPROMDevice),
     .class_init    = smbus_eeprom_class_initfn,
--
Quoted 1 line

2.17.1

thanks
-- PMM

#10 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 36 lines · Peter Maydell
Quoted 18 lines · On 7 November 2018 at 15:54, <[email protected]> wrote:

From: Corey Minyard <[email protected]>

This was if the eeprom is accessed during a state transfer, the transfer will be reliable.

Cc Paolo Bonzini <[email protected]>
Cc Michael S. Tsirkin <[email protected]>
Cc Dr. David Alan Gilbert <[email protected]>
---
  hw/i2c/smbus_eeprom.c | 16 +++++++++++++++-
  1 file changed, 15 insertions(+), 1 deletion(-)
hw/i2c/smbus_eeprom.c +0 -0
index f18aa3de35..d4430b0c5d 100644

hw/i2c/smbus_eeprom.c line 32

Quoted 4 lines
hw/i2c/smbus_eeprom.c +1 -0
@@ -29,6 +29,8 @@
 
  //#define DEBUG
 
+#define TYPE_SMBUS_EEPROM_DEVICE "smbus-eeprom"

The part of this patch which is adding the usual QOM macros is good, but we should provide also the cast-macro (the one that's an OBJECT_CHECK(...)). And this should be separate from adding the vmstate.

Quoted 8 lines
hw/i2c/smbus_eeprom.c +1 -0
+
  typedef struct SMBusEEPROMDevice {
      SMBusDevice smbusdev;
      void *data;
@@ -97,6 +99,17 @@ static uint8_t eeprom_read_data(SMBusDevice *dev, uint8_t cmd, int n)
      return eeprom_receive_byte(dev);
  }
 
Quoted 8 lines

hw/i2c/smbus_eeprom.c static uint8_t eeprom_read_data(SMBusDevice *dev, uint8_t cmd, int n) · line 103

Quoted 2 lines
hw/i2c/smbus_eeprom.c +2 -0
+static const VMStateDescription vmstate_smbus_eeprom = {
+    .name = TYPE_SMBUS_EEPROM_DEVICE,

Generally we don't use the TYPE_FOO constant for the vmstate name, because we can usually freely change the TYPE_FOO string without breaking back-compat, but we can't change the vmstate name. So we decouple them to make it more obvious when a change might be changing the migration on-the-wire format.

Ok, I've updated the code to use the standard type name and cast method (in a separate patch) and I fixed the vmstate name.  It is better, you are right.

Quoted 8 lines

hw/i2c/smbus_eeprom.c static uint8_t eeprom_read_data(SMBusDevice *dev, uint8_t cmd, int n) · line 111

Quoted 8 lines
hw/i2c/smbus_eeprom.c +8 -0
+    .version_id = 1,
+    .minimum_version_id = 1,
+    .fields      = (VMStateField[]) {
+        VMSTATE_SMBUS_DEVICE(smbusdev, SMBusEEPROMDevice),
+        VMSTATE_UINT8(offset, SMBusEEPROMDevice),
+        VMSTATE_END_OF_LIST()
+    }
+};
Quoted 10 lines

This doesn't do anything for migration of the actual data contents. The current users of this device (hw/arm/aspeed.c and the smbus_eeprom_init() function) doesn't do anything
to migrate the contents. For that matter, "user of the device passes a pointer to a random lump of memory via a device property" is a bit funky as an interface. The device should allocate its own memory and migrate it, and the user should just pass the initial required contents (default being "zero-initialize", which is what everybody except the mips_fulong2e, mips_malta and sam460ex want).

I debated on this, and it depends on what the eeprom is used for.  If it's a DRAM eeprom, it shouldn't need to be transferred. If it's something where the host stores data for later use, you do. Since it wasn't being cloned before, it probably won't matter now.

But even in the DRAM eeprom case, it shouldn't matter if it gets transferred.  So it's probably safe to just transfer it.  Easy enough to add.

Quoted 5 lines

Does this also break migration from an old QEMU to a new one? (For Aspeed that's probably ok, but we should flag it up in the commit message if so. x86 uses need care...)

There is no transfer before, so I don't see why it would break anything. Am I missing something?

-corey

Quoted 23 lines
Quoted 20 lines
hw/i2c/smbus_eeprom.c +3 -2
+
  static void smbus_eeprom_realize(DeviceState *dev, Error **errp)
  {
      SMBusEEPROMDevice *eeprom = (SMBusEEPROMDevice *)dev;
@@ -121,12 +134,13 @@ static void smbus_eeprom_class_initfn(ObjectClass *klass, void *data)
      sc->write_data = eeprom_write_data;
      sc->read_data = eeprom_read_data;
      dc->props = smbus_eeprom_properties;
+    dc->vmsd = &vmstate_smbus_eeprom;
      /* Reason: pointer property "data" */
      dc->user_creatable = false;
  }
 
  static const TypeInfo smbus_eeprom_info = {
-    .name          = "smbus-eeprom",
+    .name          = TYPE_SMBUS_EEPROM_DEVICE,
      .parent        = TYPE_SMBUS_DEVICE,
      .instance_size = sizeof(SMBusEEPROMDevice),
      .class_init    = smbus_eeprom_class_initfn,
--
Quoted 1 line

2.17.1

thanks
-- PMM

#11 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 12 lines · Corey Minyard
Quoted 10 lines · Peter Maydell

This doesn't do anything for migration of the actual data contents. The current users of this device (hw/arm/aspeed.c and the smbus_eeprom_init() function) doesn't do anything
to migrate the contents. For that matter, "user of the device passes a pointer to a random lump of memory via a device property" is a bit funky as an interface. The device should allocate its own memory and migrate it, and the user should just pass the initial required contents (default being "zero-initialize", which is what everybody except the mips_fulong2e, mips_malta and sam460ex want).

Quoted 2 lines

I debated on this, and it depends on what the eeprom is used for. If it's a DRAM eeprom, it shouldn't need to be transferred.

It's guest-visible data; the guest can write it and read it back. So it needs to be migrated. Otherwise behaviour after migration will not be the same as if the guest had never migrated.

Quoted 6 lines
Quoted 4 lines

Does this also break migration from an old QEMU to a new one? (For Aspeed that's probably ok, but we should flag it up in the commit message if so. x86 uses need care...)

There is no transfer before, so I don't see why it would break anything. Am I missing something?

I forget what the behaviour is where the source QEMU didn't have a vmstate at all but the destination QEMU does expect one. David can remind me...

thanks
- PMM

#12 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 12 lines · Peter Maydell
Quoted 11 lines · Corey Minyard
Quoted 10 lines · Peter Maydell

This doesn't do anything for migration of the actual data contents. The current users of this device (hw/arm/aspeed.c and the smbus_eeprom_init() function) doesn't do anything
to migrate the contents. For that matter, "user of the device passes a pointer to a random lump of memory via a device property" is a bit funky as an interface. The device should allocate its own memory and migrate it, and the user should just pass the initial required contents (default being "zero-initialize", which is what everybody except the mips_fulong2e, mips_malta and sam460ex want).

Quoted 5 lines
Quoted 2 lines

I debated on this, and it depends on what the eeprom is used for. If it's a DRAM eeprom, it shouldn't need to be transferred.

It's guest-visible data; the guest can write it and read it back. So it needs to be migrated. Otherwise behaviour after migration will not be the same as if the guest had never migrated.

I looked at adding it, but I ran into an issue.  The value is a

DEFINE_PROP_PTR("data", SMBusEEPROMDevice, data)

and that means the data has to be void *, but to transfer it it must be a uint8_t *. The pointer property seems to be a cheap hack, I'm not sure what it will take to fix it.

I was hoping this would be easy.  I guess transferring the eeprom is not that important, the state of pm_smbus is a lot more critical.  But it would be nice to fix it since I am messing with that code.

Thanks for your help.

-corey

Quoted 12 lines
Quoted 6 lines
Quoted 4 lines

Does this also break migration from an old QEMU to a new one? (For Aspeed that's probably ok, but we should flag it up in the commit message if so. x86 uses need care...)

There is no transfer before, so I don't see why it would break anything. Am I missing something?

I forget what the behaviour is where the source QEMU didn't have a vmstate at all but the destination QEMU does expect one. David can remind me...

thanks
- PMM

#13 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 28 lines · Corey Minyard
Quoted 16 lines · Peter Maydell
Quoted 14 lines · Corey Minyard
Quoted 11 lines · Peter Maydell

This doesn't do anything for migration of the actual data contents. The current users of this device (hw/arm/aspeed.c and the smbus_eeprom_init() function) doesn't do anything
to migrate the contents. For that matter, "user of the device passes a pointer to a random lump of memory via a device property" is a bit funky as an interface. The device should allocate its own memory and migrate it, and the user should just pass the initial required contents (default being "zero-initialize", which is what everybody except the mips_fulong2e, mips_malta and sam460ex want).

Quoted 6 lines
Quoted 2 lines

I debated on this, and it depends on what the eeprom is used for. If it's a DRAM eeprom, it shouldn't need to be transferred.

It's guest-visible data; the guest can write it and read it back. So it needs to be migrated. Otherwise behaviour after migration will not be the same as if the guest had never migrated.

I looked at adding it, but I ran into an issue. The value is a

Quoted 7 lines

DEFINE_PROP_PTR("data", SMBusEEPROMDevice, data)

and that means the data has to be void *, but to transfer it it must be a uint8_t *.
The pointer property seems to be a cheap hack, I'm not sure what it will take
to fix it.

The data provided by the caller is only the initialization data. So I think the device should create its own memory to copy that init data into, and migrate that ram, not wherever the initialization data lives. (Currently this "copy the data into our own ram" happens in the smbus_eeprom_init() wrapper, but we should do it in the device realize function instead.)

Also there seem to be unresolved questions about what happens on reset -- should the EEPROM revert back to the initial contents? We don't do that at the moment, but this breaks the usual QEMU pattern that reset is as if you'd just completely restarted QEMU.

thanks
-- PMM

#14 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 30 lines · Peter Maydell
Quoted 22 lines · Corey Minyard
Quoted 12 lines · Peter Maydell
Quoted 11 lines · Corey Minyard
Quoted 10 lines · Peter Maydell

This doesn't do anything for migration of the actual data contents. The current users of this device (hw/arm/aspeed.c and the smbus_eeprom_init() function) doesn't do anything
to migrate the contents. For that matter, "user of the device passes a pointer to a random lump of memory via a device property" is a bit funky as an interface. The device should allocate its own memory and migrate it, and the user should just pass the initial required contents (default being "zero-initialize", which is what everybody except the mips_fulong2e, mips_malta and sam460ex want).

Quoted 5 lines
Quoted 2 lines

I debated on this, and it depends on what the eeprom is used for. If it's a DRAM eeprom, it shouldn't need to be transferred.

It's guest-visible data; the guest can write it and read it back. So it needs to be migrated. Otherwise behaviour after migration will not be the same as if the guest had never migrated.

I looked at adding it, but I ran into an issue. The value is a

Quoted 7 lines

DEFINE_PROP_PTR("data", SMBusEEPROMDevice, data)

and that means the data has to be void *, but to transfer it it must be a uint8_t *.
The pointer property seems to be a cheap hack, I'm not sure what it will take
to fix it.

Quoted 7 lines

The data provided by the caller is only the initialization data. So I think the device should create its own memory to copy that init data into, and migrate that ram, not wherever the initialization data lives. (Currently this "copy the data into our own ram" happens in the smbus_eeprom_init() wrapper, but we should do it in the device realize function instead.)

That's what I would like, but should I get rid of the "DEFINE_PROP_PTR"? I don't know the value of creating a properly like this, since the user can't set it and can't see it.  Plus the comments in the code say that it should be gotten rid of at some point.

But if I store off the initialization data and keep the actual data in a separate memory created by the realize function, that should work find.

Quoted 6 lines

Also there seem to be unresolved questions about what happens on reset -- should the EEPROM revert back to the initial contents? We don't do that at the moment, but this breaks the usual QEMU pattern that reset is as if you'd just completely restarted QEMU.

I would consider this part of the hardware, like data on a disk drive, so it seem better to leave it alone after a reset.  But it's not quite the same.  So I'm not sure.

Thanks,

-corey

Quoted 2 lines

thanks
-- PMM

#15 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 10 lines · Corey Minyard
Quoted 7 lines · Peter Maydell

The data provided by the caller is only the initialization data. So I think the device should create its own memory to copy that init data into, and migrate that ram, not wherever the initialization data lives. (Currently this "copy the data into our own ram" happens in the smbus_eeprom_init() wrapper, but we should do it in the device realize function instead.)

Quoted 8 lines

That's what I would like, but should I get rid of the "DEFINE_PROP_PTR"? I don't know the value of creating a properly like this, since the user can't set it and can't see it. Plus the comments in the code say that it should be gotten rid of at some point.

But if I store off the initialization data and keep the actual data in a separate memory created by the realize function, that should work find.

Well, you still have to pass the init data to the device somehow, so I think a pointer property is not a terrible way to do that.

Quoted 10 lines
Quoted 5 lines

Also there seem to be unresolved questions about what happens on reset -- should the EEPROM revert back to the initial contents? We don't do that at the moment, but this breaks the usual QEMU pattern that reset is as if you'd just completely restarted QEMU.

I would consider this part of the hardware, like data on a disk drive, so it seem better to leave it alone after a reset. But it's not quite the same. So I'm not sure.

That would require us to support backing it properly with a block device, like the pflash flash devices, I think. (This would be the long term way to be able to dump the pointer property, in favour of a block backend property.)

thanks
-- PMM

#16 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 31 lines · * Peter Maydell ([email protected]) wrote:
Quoted 10 lines · Corey Minyard
Quoted 7 lines · Peter Maydell

The data provided by the caller is only the initialization data. So I think the device should create its own memory to copy that init data into, and migrate that ram, not wherever the initialization data lives. (Currently this "copy the data into our own ram" happens in the smbus_eeprom_init() wrapper, but we should do it in the device realize function instead.)

Quoted 8 lines

That's what I would like, but should I get rid of the "DEFINE_PROP_PTR"? I don't know the value of creating a properly like this, since the user can't set it and can't see it. Plus the comments in the code say that it should be gotten rid of at some point.

But if I store off the initialization data and keep the actual data in a separate memory created by the realize function, that should work find.

Well, you still have to pass the init data to the device somehow, so I think a pointer property is not a terrible way to do that.

Quoted 7 lines
Quoted 5 lines

Also there seem to be unresolved questions about what happens on reset -- should the EEPROM revert back to the initial contents? We don't do that at the moment, but this breaks the usual QEMU pattern that reset is as if you'd just completely restarted QEMU.

Quoted 8 lines
Quoted 3 lines

I would consider this part of the hardware, like data on a disk drive, so it seem better to leave it alone after a reset. But it's not quite the same. So I'm not sure.

That would require us to support backing it properly with a block device, like the pflash flash devices, I think. (This would be the long term way to be able to dump the pointer property, in favour of a block backend property.)

There's a number of separate questions:

a) Can the guest write to the EEPROM or are we just treating it as a ROM?
b) If a guest writes to the EEPROM and then resets does the data stay there [I'd expect so, it's an EEPROM]
c) It the guest writes to the EEPROM and then quits qemu and restarts does the data stay there?
d) If the guest is migrated does it keep the data it's written - I'd say it must because otherwise it would get confused.

(c) is where it starts looking like a pflash - which does get a bit messy.

Dave

Quoted 2 lines

thanks
-- PMM

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK

#17 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

Quoted 17 lines · Dr. David Alan Gilbert
Quoted 4 lines · * Peter Maydell ([email protected]) wrote:

That would require us to support backing it properly with a block device, like the pflash flash devices, I think. (This would be the long term way to be able to dump the pointer property, in favour of a block backend property.)

There's a number of separate questions:

a) Can the guest write to the EEPROM or are we just treating it as a ROM?
b) If a guest writes to the EEPROM and then resets does the data stay there [I'd expect so, it's an EEPROM]
c) It the guest writes to the EEPROM and then quits qemu and restarts does the data stay there?
d) If the guest is migrated does it keep the data it's written - I'd say it must because otherwise it would get confused.

Quoted 2 lines

(c) is where it starts looking like a pflash - which does get a bit messy.

My view is that the answers to b and c should always be the same:
resetting QEMU with a system reset should look the same as if you stopped QEMU and restarted it. (Our only kind of system reset is an emulation of a hard powercycle.)

In this case, the answer to "a" is currently "yes, it can write to it", though whether any guest takes advantage of that is a different question.

thanks
-- PMM

#18 lore raw

Re: [Qemu-devel] [PATCH 3/3] i2c: Add vmstate handling to the smbus eeprom

In reply to Peter Maydell
Quoted 24 lines · * Peter Maydell ([email protected]) wrote:
Quoted 12 lines · Corey Minyard
Quoted 10 lines · Peter Maydell

This doesn't do anything for migration of the actual data contents. The current users of this device (hw/arm/aspeed.c and the smbus_eeprom_init() function) doesn't do anything
to migrate the contents. For that matter, "user of the device passes a pointer to a random lump of memory via a device property" is a bit funky as an interface. The device should allocate its own memory and migrate it, and the user should just pass the initial required contents (default being "zero-initialize", which is what everybody except the mips_fulong2e, mips_malta and sam460ex want).

Quoted 2 lines

I debated on this, and it depends on what the eeprom is used for. If it's a DRAM eeprom, it shouldn't need to be transferred.

It's guest-visible data; the guest can write it and read it back. So it needs to be migrated. Otherwise behaviour after migration will not be the same as if the guest had never migrated.

Quoted 4 lines
Quoted 4 lines

Does this also break migration from an old QEMU to a new one? (For Aspeed that's probably ok, but we should flag it up in the commit message if so. x86 uses need care...)

Quoted 6 lines
Quoted 2 lines

There is no transfer before, so I don't see why it would break anything. Am I missing something?

I forget what the behaviour is where the source QEMU didn't have a vmstate at all but the destination QEMU does expect one. David can remind me...

If it's a separate device then you tend to get away with it - there's no check that all devices receive their state, so it should work. This does break backwards migration though - migrating to an older qemu with the existing machine type will complain it's received a device which it doesn't understand.
You should be able to add a .needed to the device
so it's only sent for new machine types.
(Which is what I said in August, when I also asked about the data)

Dave

Quoted 2 lines

thanks
- PMM

--
Dr. David Alan Gilbert / [email protected] / Manchester, UK