Skip to contentSkip to conversation

PCDIMM cleanup

PATCH In review Wei Yang 6 participants 22 messages Feb 20, 2019 – Mar 1, 2019

lore
Share & tools
lore

Wei Yang

author
#0 lore raw

v2 cover PCDIMM cleanup

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

v2:
* fix warning in Patch 1
* split Patch 2 into two

Wei Yang (3):
pc-dimm: remove check on pc-dimm hotpluggable
mem/nvdimm: remove nvdimm_realize
pc-dimm: revert "introduce realize callback"

 hw/acpi/memory_hotplug.c |  5 -----
 hw/mem/nvdimm.c          | 11 -----------
 hw/mem/pc-dimm.c         |  5 -----
 include/hw/mem/pc-dimm.h |  3 ---
 4 files changed, 24 deletions(-)
Signature · 1 line
2.19.1
author
#1 lore raw

PATCH v2 1/3 pc-dimm: remove check on pc-dimm hotpluggable

Function acpi_memory_plug_cb() is only invoked when dev is a PCDIMM, which is hotpluggable. This means it is not necessary to check this property again.

This patch removes this check.

---
v2:
* remove unused dc

---
 hw/acpi/memory_hotplug.c | 5 -----
 1 file changed, 5 deletions(-)
hw/acpi/memory_hotplug.c +0 -5
index 8c7c1013f3..cb5284d36f 100644
@@ -264,11 +264,6 @@ void acpi_memory_plug_cb(HotplugHandler *hotplug_dev, MemHotplugState *mem_st,
                          DeviceState *dev, Error **errp)
 {
     MemStatus *mdev;
-    DeviceClass *dc = DEVICE_GET_CLASS(dev);
-
-    if (!dc->hotpluggable) {
-        return;
-    }
 
     mdev = acpi_memory_slot_status(mem_st, dev, errp);
     if (!mdev) {
Signature · 1 line
2.19.1
Tested-by on v2 1/3
#2 lore raw

Re: [Qemu-devel] [PATCH v2 1/3] pc-dimm: remove check on pc-dimm hotpluggable

Quoted 8 lines · Wei Yang

Function acpi_memory_plug_cb() is only invoked when dev is a PCDIMM, which is hotpluggable. This means it is not necessary to check this property again.

This patch removes this check.

Quoted 24 lines

---
v2:
* remove unused dc

---
 hw/acpi/memory_hotplug.c | 5 -----
 1 file changed, 5 deletions(-)
hw/acpi/memory_hotplug.c +0 -5
index 8c7c1013f3..cb5284d36f 100644
@@ -264,11 +264,6 @@ void acpi_memory_plug_cb(HotplugHandler *hotplug_dev, MemHotplugState *mem_st,
                          DeviceState *dev, Error **errp)
 {
     MemStatus *mdev;
-    DeviceClass *dc = DEVICE_GET_CLASS(dev);
-
-    if (!dc->hotpluggable) {
-        return;
-    }
 
     mdev = acpi_memory_slot_status(mem_st, dev, errp);
     if (!mdev) {
author
#3 lore raw

PATCH v2 2/3 mem/nvdimm: remove nvdimm_realize

In reply to Wei Yang

nvdimm_realize is used to prepare its memory region, while this is done in pre_plug stage.

This is time to remove it.

---
v2: split nvdimm part here

---
 hw/mem/nvdimm.c | 11 -----------
 1 file changed, 11 deletions(-)
hw/mem/nvdimm.c +0 -11
index bf2adf5e16..8f69576926 100644
@@ -136,15 +136,6 @@ static MemoryRegion *nvdimm_md_get_memory_region(MemoryDeviceState *md,
     return nvdimm->nvdimm_mr;
 }
 
-static void nvdimm_realize(PCDIMMDevice *dimm, Error **errp)
-{
-    NVDIMMDevice *nvdimm = NVDIMM(dimm);
-
-    if (!nvdimm->nvdimm_mr) {
-        nvdimm_prepare_memory_region(nvdimm, errp);
-    }
-}
-
 /*
  * the caller should check the input parameters before calling
  * label read/write functions.
@@ -192,12 +183,10 @@ static Property nvdimm_properties[] = {
 
 static void nvdimm_class_init(ObjectClass *oc, void *data)
 {
-    PCDIMMDeviceClass *ddc = PC_DIMM_CLASS(oc);
     MemoryDeviceClass *mdc = MEMORY_DEVICE_CLASS(oc);
     NVDIMMClass *nvc = NVDIMM_CLASS(oc);
     DeviceClass *dc = DEVICE_CLASS(oc);
 
-    ddc->realize = nvdimm_realize;
     mdc->get_memory_region = nvdimm_md_get_memory_region;
     dc->props = nvdimm_properties;
 
Signature · 1 line
2.19.1
Reviewed-by on v2 2/3
#4 lore raw

Re: [Qemu-devel] [PATCH v2 2/3] mem/nvdimm: remove nvdimm_realize

Quoted 2 lines · Wei Yang

nvdimm_realize is used to prepare its memory region, while this is done in pre_plug stage.

Via the device parent:

pc_dimm_pre_plug()
-> memory_device_pre_plug()

     -> MemoryDeviceClass::get_memory_region()
        nvdimm_md_get_memory_region()
Quoted 4 lines

This is time to remove it.

Quoted 41 lines

---
v2: split nvdimm part here

---
 hw/mem/nvdimm.c | 11 -----------
 1 file changed, 11 deletions(-)
hw/mem/nvdimm.c +0 -11
index bf2adf5e16..8f69576926 100644
@@ -136,15 +136,6 @@ static MemoryRegion *nvdimm_md_get_memory_region(MemoryDeviceState *md,
     return nvdimm->nvdimm_mr;
 }
 
-static void nvdimm_realize(PCDIMMDevice *dimm, Error **errp)
-{
-    NVDIMMDevice *nvdimm = NVDIMM(dimm);
-
-    if (!nvdimm->nvdimm_mr) {
-        nvdimm_prepare_memory_region(nvdimm, errp);
-    }
-}
-
 /*
  * the caller should check the input parameters before calling
  * label read/write functions.
@@ -192,12 +183,10 @@ static Property nvdimm_properties[] = {
 
 static void nvdimm_class_init(ObjectClass *oc, void *data)
 {
-    PCDIMMDeviceClass *ddc = PC_DIMM_CLASS(oc);
     MemoryDeviceClass *mdc = MEMORY_DEVICE_CLASS(oc);
     NVDIMMClass *nvc = NVDIMM_CLASS(oc);
     DeviceClass *dc = DEVICE_CLASS(oc);
 
-    ddc->realize = nvdimm_realize;
     mdc->get_memory_region = nvdimm_md_get_memory_region;
     dc->props = nvdimm_properties;
 
author
#5 lore raw

PATCH v2 3/3 pc-dimm: revert "introduce realize callback"

In reply to Wei Yang

realize callback in introduced to check if the backend memory is large enough to contain label data and init its memory region, while this task is handled in pre_plug stage.

Now it's time to remove it.

---
 hw/mem/pc-dimm.c         | 5 -----
 include/hw/mem/pc-dimm.h | 3 ---
 2 files changed, 8 deletions(-)
hw/mem/pc-dimm.c +0 -5
index 152400b1fc..5832c0ba92 100644
@@ -159,7 +159,6 @@ static void pc_dimm_init(Object *obj)
 static void pc_dimm_realize(DeviceState *dev, Error **errp)
 {
     PCDIMMDevice *dimm = PC_DIMM(dev);
-    PCDIMMDeviceClass *ddc = PC_DIMM_GET_CLASS(dimm);
 
     if (!dimm->hostmem) {
         error_setg(errp, "'" PC_DIMM_MEMDEV_PROP "' property is not set");
@@ -178,10 +177,6 @@ static void pc_dimm_realize(DeviceState *dev, Error **errp)
         return;
     }
 
-    if (ddc->realize) {
-        ddc->realize(dimm, errp);
-    }
-
     host_memory_backend_set_mapped(dimm->hostmem, true);
 }
 
include/hw/mem/pc-dimm.h +0 -3
index 01436b9f50..d18f8246b7 100644
@@ -59,8 +59,6 @@ typedef struct PCDIMMDevice {
 
 /**
  * PCDIMMDeviceClass:
- * @realize: called after common dimm is realized so that the dimm based
- * devices get the chance to do specified operations.
  * @get_vmstate_memory_region: returns #MemoryRegion which indicates the
  * memory of @dimm should be kept during live migration. Will not fail
  * after the device was realized.
@@ -70,7 +68,6 @@ typedef struct PCDIMMDeviceClass {
     DeviceClass parent_class;
 
     /* public */
-    void (*realize)(PCDIMMDevice *dimm, Error **errp);
     MemoryRegion *(*get_vmstate_memory_region)(PCDIMMDevice *dimm,
                                                Error **errp);
 } PCDIMMDeviceClass;
Signature · 1 line
2.19.1
Reviewed-by on v2 3/3
#6 lore raw

Re: [Qemu-devel] [PATCH v2 3/3] pc-dimm: revert "introduce realize callback"

Quoted 5 lines · Wei Yang

realize callback in introduced to check if the backend memory is large enough to contain label data and init its memory region, while this task is handled in pre_plug stage.

Now it's time to remove it.

Good cleanup!

Michael, can you add:

"This reverts commit 9f318f8f7e689b9653b42bac73047f9719a1f34e."

Thanks,

Phil.

Quoted 2 lines
Quoted 50 lines
---
 hw/mem/pc-dimm.c         | 5 -----
 include/hw/mem/pc-dimm.h | 3 ---
 2 files changed, 8 deletions(-)
hw/mem/pc-dimm.c +0 -5
index 152400b1fc..5832c0ba92 100644
@@ -159,7 +159,6 @@ static void pc_dimm_init(Object *obj)
 static void pc_dimm_realize(DeviceState *dev, Error **errp)
 {
     PCDIMMDevice *dimm = PC_DIMM(dev);
-    PCDIMMDeviceClass *ddc = PC_DIMM_GET_CLASS(dimm);
 
     if (!dimm->hostmem) {
         error_setg(errp, "'" PC_DIMM_MEMDEV_PROP "' property is not set");
@@ -178,10 +177,6 @@ static void pc_dimm_realize(DeviceState *dev, Error **errp)
         return;
     }
 
-    if (ddc->realize) {
-        ddc->realize(dimm, errp);
-    }
-
     host_memory_backend_set_mapped(dimm->hostmem, true);
 }
 
include/hw/mem/pc-dimm.h +0 -3
index 01436b9f50..d18f8246b7 100644
@@ -59,8 +59,6 @@ typedef struct PCDIMMDevice {
 
 /**
  * PCDIMMDeviceClass:
- * @realize: called after common dimm is realized so that the dimm based
- * devices get the chance to do specified operations.
  * @get_vmstate_memory_region: returns #MemoryRegion which indicates the
  * memory of @dimm should be kept during live migration. Will not fail
  * after the device was realized.
@@ -70,7 +68,6 @@ typedef struct PCDIMMDeviceClass {
     DeviceClass parent_class;
 
     /* public */
-    void (*realize)(PCDIMMDevice *dimm, Error **errp);
     MemoryRegion *(*get_vmstate_memory_region)(PCDIMMDevice *dimm,
                                                Error **errp);
 } PCDIMMDeviceClass;
author
#7 lore raw

Re: [Qemu-devel] [PATCH v2 3/3] pc-dimm: revert "introduce realize callback"

Quoted 9 lines · Philippe Mathieu-Daudé
Quoted 5 lines · Wei Yang

realize callback in introduced to check if the backend memory is large enough to contain label data and init its memory region, while this task is handled in pre_plug stage.

Now it's time to remove it.

Good cleanup!

Glad you like it :-)

Quoted 63 lines

Michael, can you add:

"This reverts commit 9f318f8f7e689b9653b42bac73047f9719a1f34e."

Thanks,

Phil.

Quoted 2 lines
Quoted 50 lines
---
 hw/mem/pc-dimm.c         | 5 -----
 include/hw/mem/pc-dimm.h | 3 ---
 2 files changed, 8 deletions(-)
hw/mem/pc-dimm.c +0 -5
index 152400b1fc..5832c0ba92 100644
@@ -159,7 +159,6 @@ static void pc_dimm_init(Object *obj)
 static void pc_dimm_realize(DeviceState *dev, Error **errp)
 {
     PCDIMMDevice *dimm = PC_DIMM(dev);
-    PCDIMMDeviceClass *ddc = PC_DIMM_GET_CLASS(dimm);
 
     if (!dimm->hostmem) {
         error_setg(errp, "'" PC_DIMM_MEMDEV_PROP "' property is not set");
@@ -178,10 +177,6 @@ static void pc_dimm_realize(DeviceState *dev, Error **errp)
         return;
     }
 
-    if (ddc->realize) {
-        ddc->realize(dimm, errp);
-    }
-
     host_memory_backend_set_mapped(dimm->hostmem, true);
 }
 
include/hw/mem/pc-dimm.h +0 -3
index 01436b9f50..d18f8246b7 100644
@@ -59,8 +59,6 @@ typedef struct PCDIMMDevice {
 
 /**
  * PCDIMMDeviceClass:
- * @realize: called after common dimm is realized so that the dimm based
- * devices get the chance to do specified operations.
  * @get_vmstate_memory_region: returns #MemoryRegion which indicates the
  * memory of @dimm should be kept during live migration. Will not fail
  * after the device was realized.
@@ -70,7 +68,6 @@ typedef struct PCDIMMDeviceClass {
     DeviceClass parent_class;
 
     /* public */
-    void (*realize)(PCDIMMDevice *dimm, Error **errp);
     MemoryRegion *(*get_vmstate_memory_region)(PCDIMMDevice *dimm,
                                                Error **errp);
 } PCDIMMDeviceClass;
Signature · 2 lines
Wei Yang
Help you, Help me
Reviewed-by on the v2 series
#8 lore raw
In reply to Wei Yang
Quoted 11 lines · Wei Yang

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

v2:
* fix warning in Patch 1
* split Patch 2 into two

Quoted 5 lines

Wei Yang (3):
pc-dimm: remove check on pc-dimm hotpluggable
mem/nvdimm: remove nvdimm_realize

pc-dimm: revert "introduce realize callback"

I think the word 'revert' is not so precise as that hints the commit is bugly, instead, it was factored in the later comments then becomes useless now.

Anyway, this pathset looks good to me.

author
#9 lore raw
Quoted 18 lines · Xiao Guangrong
Quoted 14 lines · Wei Yang

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

v2:
* fix warning in Patch 1
* split Patch 2 into two

Wei Yang (3):
pc-dimm: remove check on pc-dimm hotpluggable
mem/nvdimm: remove nvdimm_realize

Quoted 6 lines
Quoted 1 line

pc-dimm: revert "introduce realize callback"

I think the word 'revert' is not so precise as that hints the commit is bugly, instead, it was factored in the later comments then becomes useless now.

You are right. It is always difficult for me to pick up the proper word.

Quoted 3 lines

Anyway, this pathset looks good to me.

Thanks, Xiao.

Signature · 2 lines
Wei Yang
Help you, Help me
#10 lore raw
In reply to Wei Yang
Quoted 6 lines · On Wed, 20 Feb 2019 08:51:21 +0800 Wei Yang <[email protected]>…

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

even though this series doesn't break anything, I disagree with it conceptually as it makes device less abstracted and make it more dependent on how existing machine code uses it.
I'd drop whole series.

Quoted 16 lines

v2:
* fix warning in Patch 1
* split Patch 2 into two

Wei Yang (3):
pc-dimm: remove check on pc-dimm hotpluggable
mem/nvdimm: remove nvdimm_realize
pc-dimm: revert "introduce realize callback"

 hw/acpi/memory_hotplug.c |  5 -----
 hw/mem/nvdimm.c          | 11 -----------
 hw/mem/pc-dimm.c         |  5 -----
 include/hw/mem/pc-dimm.h |  3 ---
 4 files changed, 24 deletions(-)
#11 lore raw
Quoted 5 lines · Igor Mammedov
Quoted 2 lines · On Wed, 20 Feb 2019 08:51:21 +0800 Wei Yang <[email protected]>…

Three trivial cleanup for pc-dimm.

Quoted 9 lines
Quoted 4 lines

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

even though this series doesn't break anything, I disagree with it conceptually as it makes device less abstracted and make it more dependent on how existing machine code uses it.
I'd drop whole series.

Is Patch [1] also make device more dependent on existing implementation?

For example, when we look at the counterpart of acpi_memory_plug_cb():

acpi_pcihp_device_plug_cb

which handle the pci device hotplug. We don't check the hotpluggable property for pci devices.

To me, this is a general rule for PCDIMM, they are hotpluggable.

For Patch[2][3], I agree with you.

Signature · 2 lines
Wei Yang
Help you, Help me
#12 lore raw
Quoted 10 lines · On Sat, 23 Feb 2019 00:02:49 +0000 Wei Yang <[email protected]> wrote:
Quoted 9 lines · Igor Mammedov
Quoted 6 lines · On Wed, 20 Feb 2019 08:51:21 +0800 Wei Yang <[email protected]>…

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

Quoted 7 lines
Quoted 5 lines

even though this series doesn't break anything, I disagree with it conceptually as it makes device less abstracted and make it more dependent on how existing machine code uses it.
I'd drop whole series.

Is Patch [1] also make device more dependent on existing implementation?

yes, it's implementation detail that PCDIMM&Co are hotpluggable now.

Quoted 8 lines

For example, when we look at the counterpart of acpi_memory_plug_cb():

acpi_pcihp_device_plug_cb

which handle the pci device hotplug. We don't check the hotpluggable property for pci devices.

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Quoted 2 lines

For Patch[2][3], I agree with you.

#13 lore raw
Quoted 29 lines · Igor Mammedov
Quoted 10 lines · On Sat, 23 Feb 2019 00:02:49 +0000 Wei Yang <[email protected]> wrote:
Quoted 9 lines · Igor Mammedov
Quoted 6 lines · On Wed, 20 Feb 2019 08:51:21 +0800 Wei Yang <[email protected]>…

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

Quoted 7 lines
Quoted 5 lines

even though this series doesn't break anything, I disagree with it conceptually as it makes device less abstracted and make it more dependent on how existing machine code uses it.
I'd drop whole series.

Is Patch [1] also make device more dependent on existing implementation?

yes, it's implementation detail that PCDIMM&Co are hotpluggable now.

Quoted 7 lines

For example, when we look at the counterpart of acpi_memory_plug_cb():

acpi_pcihp_device_plug_cb

which handle the pci device hotplug. We don't check the hotpluggable property for pci devices.

Quoted 7 lines
Quoted 1 line

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Ok, if we'd support un-hotpluggable mem device, we could retain this check. But this maybe not a proper place.

Based on my understanding, at this point, every thing has been done properly. If this check pass, we will send an acpi interrupt to notify the guest. In case this is an un-hotpluggable device, every thing looks good but no effect in guest. Because we skip the notification.

Maybe we need to move the check in pre-plug?

Quoted 2 lines
Quoted 2 lines

For Patch[2][3], I agree with you.

Signature · 2 lines
Wei Yang
Help you, Help me
#14 lore raw
Quoted 41 lines · On Mon, 25 Feb 2019 12:47:14 +0000 Wei Yang <[email protected]> wrote:
Quoted 29 lines · Igor Mammedov
Quoted 10 lines · On Sat, 23 Feb 2019 00:02:49 +0000 Wei Yang <[email protected]> wrote:
Quoted 9 lines · Igor Mammedov
Quoted 6 lines · On Wed, 20 Feb 2019 08:51:21 +0800 Wei Yang <[email protected]>…

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

Quoted 7 lines
Quoted 5 lines

even though this series doesn't break anything, I disagree with it conceptually as it makes device less abstracted and make it more dependent on how existing machine code uses it.
I'd drop whole series.

Is Patch [1] also make device more dependent on existing implementation?

yes, it's implementation detail that PCDIMM&Co are hotpluggable now.

Quoted 7 lines

For example, when we look at the counterpart of acpi_memory_plug_cb():

acpi_pcihp_device_plug_cb

which handle the pci device hotplug. We don't check the hotpluggable property for pci devices.

Quoted 7 lines
Quoted 1 line

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Ok, if we'd support un-hotpluggable mem device, we could retain this check. But this maybe not a proper place.

Quoted 6 lines

Based on my understanding, at this point, every thing has been done properly. If this check pass, we will send an acpi interrupt to notify the guest. In case this is an un-hotpluggable device, every thing looks good but no effect in guest. Because we skip the notification.

Maybe we need to move the check in pre-plug?

And what would happen then and afterwards?

Point is to make one of the handlers in chain to ignore plug event (i.e. do not generate SCI event) while the rest of handlers complete successfully mapping device in address space and whatever else.

Quoted 3 lines
Quoted 2 lines
Quoted 2 lines

For Patch[2][3], I agree with you.

#15 lore raw
Quoted 49 lines · Igor Mammedov
Quoted 46 lines · On Mon, 25 Feb 2019 12:47:14 +0000 Wei Yang <[email protected]> wrote:
Quoted 29 lines · Igor Mammedov
Quoted 10 lines · On Sat, 23 Feb 2019 00:02:49 +0000 Wei Yang <[email protected]> wrote:
Quoted 9 lines · Igor Mammedov
Quoted 6 lines · On Wed, 20 Feb 2019 08:51:21 +0800 Wei Yang <[email protected]>…

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

Quoted 7 lines
Quoted 5 lines

even though this series doesn't break anything, I disagree with it conceptually as it makes device less abstracted and make it more dependent on how existing machine code uses it.
I'd drop whole series.

Is Patch [1] also make device more dependent on existing implementation?

yes, it's implementation detail that PCDIMM&Co are hotpluggable now.

Quoted 7 lines

For example, when we look at the counterpart of acpi_memory_plug_cb():

acpi_pcihp_device_plug_cb

which handle the pci device hotplug. We don't check the hotpluggable property for pci devices.

Quoted 7 lines
Quoted 1 line

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Ok, if we'd support un-hotpluggable mem device, we could retain this check. But this maybe not a proper place.

Based on my understanding, at this point, every thing has been done properly. If this check pass, we will send an acpi interrupt to notify the guest. In case this is an un-hotpluggable device, every thing looks good but no effect in guest. Because we skip the notification.

Quoted 7 lines
Quoted 1 line

Maybe we need to move the check in pre-plug?

And what would happen then and afterwards?

Point is to make one of the handlers in chain to ignore plug event (i.e. do not generate SCI event) while the rest of handlers complete successfully mapping device in address space and whatever else.

This will have a well prepared device in system but guest is not notified, right?

My idea to move the check in pre-plug will result the qemu return when it see no SCI event will be generated, so there is no device created.

I guess current behavior is a designed decision?

Quoted 3 lines
Quoted 3 lines
Quoted 2 lines
Quoted 2 lines

For Patch[2][3], I agree with you.

Signature · 2 lines
Wei Yang
Help you, Help me
#16 lore raw
Quoted 53 lines · On Wed, 27 Feb 2019 13:59:20 +0000 Wei Yang <[email protected]> wrote:
Quoted 52 lines · Igor Mammedov
Quoted 41 lines · On Mon, 25 Feb 2019 12:47:14 +0000 Wei Yang <[email protected]> wrote:
Quoted 29 lines · Igor Mammedov
Quoted 10 lines · On Sat, 23 Feb 2019 00:02:49 +0000 Wei Yang <[email protected]> wrote:
Quoted 9 lines · Igor Mammedov
Quoted 6 lines · On Wed, 20 Feb 2019 08:51:21 +0800 Wei Yang <[email protected]>…

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

Quoted 7 lines
Quoted 5 lines

even though this series doesn't break anything, I disagree with it conceptually as it makes device less abstracted and make it more dependent on how existing machine code uses it.
I'd drop whole series.

Is Patch [1] also make device more dependent on existing implementation?

yes, it's implementation detail that PCDIMM&Co are hotpluggable now.

Quoted 7 lines

For example, when we look at the counterpart of acpi_memory_plug_cb():

acpi_pcihp_device_plug_cb

which handle the pci device hotplug. We don't check the hotpluggable property for pci devices.

Quoted 7 lines
Quoted 1 line

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Ok, if we'd support un-hotpluggable mem device, we could retain this check. But this maybe not a proper place.

Quoted 6 lines

Based on my understanding, at this point, every thing has been done properly. If this check pass, we will send an acpi interrupt to notify the guest. In case this is an un-hotpluggable device, every thing looks good but no effect in guest. Because we skip the notification.

Maybe we need to move the check in pre-plug?

And what would happen then and afterwards?

Quoted 7 lines
Quoted 4 lines

Point is to make one of the handlers in chain to ignore plug event (i.e. do not generate SCI event) while the rest of handlers complete successfully mapping device in address space and whatever else.

This will have a well prepared device in system but guest is not notified, right?

yes, it's theoretically possible to move check up the call stack to machine level and not call secondary hotplug handler on non hotplugged device but that again depends on what secondary hotplug handler does. I'd rather keep logic independent here unless there is good reason not to do so.

Quoted 4 lines

My idea to move the check in pre-plug will result the qemu return when it see no SCI event will be generated, so there is no device created.

I guess current behavior is a designed decision?

I'd say so.
PS:
QEMU's hotplug_hadler API is misnamed at this point as it handles both cold (basic device wiring) and hotplug (processing hotplug). Maybe we should rename it but I'm not irritated enough by it to do so.

Quoted 4 lines
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines
Quoted 2 lines

For Patch[2][3], I agree with you.

#17 lore raw
Quoted 73 lines · Igor Mammedov
Quoted 53 lines · On Wed, 27 Feb 2019 13:59:20 +0000 Wei Yang <[email protected]> wrote:
Quoted 52 lines · Igor Mammedov
Quoted 41 lines · On Mon, 25 Feb 2019 12:47:14 +0000 Wei Yang <[email protected]> wrote:
Quoted 29 lines · Igor Mammedov
Quoted 10 lines · On Sat, 23 Feb 2019 00:02:49 +0000 Wei Yang <[email protected]> wrote:
Quoted 9 lines · Igor Mammedov
Quoted 6 lines · On Wed, 20 Feb 2019 08:51:21 +0800 Wei Yang <[email protected]>…

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

Quoted 7 lines
Quoted 5 lines

even though this series doesn't break anything, I disagree with it conceptually as it makes device less abstracted and make it more dependent on how existing machine code uses it.
I'd drop whole series.

Is Patch [1] also make device more dependent on existing implementation?

yes, it's implementation detail that PCDIMM&Co are hotpluggable now.

Quoted 7 lines

For example, when we look at the counterpart of acpi_memory_plug_cb():

acpi_pcihp_device_plug_cb

which handle the pci device hotplug. We don't check the hotpluggable property for pci devices.

Quoted 7 lines
Quoted 1 line

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Ok, if we'd support un-hotpluggable mem device, we could retain this check. But this maybe not a proper place.

Quoted 6 lines

Based on my understanding, at this point, every thing has been done properly. If this check pass, we will send an acpi interrupt to notify the guest. In case this is an un-hotpluggable device, every thing looks good but no effect in guest. Because we skip the notification.

Maybe we need to move the check in pre-plug?

And what would happen then and afterwards?

Quoted 7 lines
Quoted 4 lines

Point is to make one of the handlers in chain to ignore plug event (i.e. do not generate SCI event) while the rest of handlers complete successfully mapping device in address space and whatever else.

This will have a well prepared device in system but guest is not notified, right?

yes, it's theoretically possible to move check up the call stack to machine level and not call secondary hotplug handler on non hotplugged device but that again depends on what secondary hotplug handler does. I'd rather keep logic independent here unless there is good reason not to do so.

Quoted 3 lines

My idea to move the check in pre-plug will result the qemu return when it see no SCI event will be generated, so there is no device created.

Quoted 7 lines
Quoted 1 line

I guess current behavior is a designed decision?

I'd say so.
PS:
QEMU's hotplug_hadler API is misnamed at this point as it handles both cold (basic device wiring) and hotplug (processing hotplug). Maybe we should rename it but I'm not irritated enough by it to do so.

Got it, thanks.

Quoted 4 lines
Quoted 4 lines
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines
Quoted 2 lines

For Patch[2][3], I agree with you.

Signature · 2 lines
Wei Yang
Help you, Help me
author
#18 lore raw
In reply to Igor Mammedov
Quoted 43 lines · Igor Mammedov
Quoted 23 lines · On Wed, 27 Feb 2019 13:59:20 +0000 Wei Yang <[email protected]> wrote:
Quoted 22 lines · Igor Mammedov
Quoted 11 lines · On Mon, 25 Feb 2019 12:47:14 +0000 Wei Yang <[email protected]> wrote:
Quoted 7 lines
Quoted 1 line

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Ok, if we'd support un-hotpluggable mem device, we could retain this check. But this maybe not a proper place.

Quoted 6 lines

Based on my understanding, at this point, every thing has been done properly. If this check pass, we will send an acpi interrupt to notify the guest. In case this is an un-hotpluggable device, every thing looks good but no effect in guest. Because we skip the notification.

Maybe we need to move the check in pre-plug?

And what would happen then and afterwards?

Quoted 7 lines
Quoted 4 lines

Point is to make one of the handlers in chain to ignore plug event (i.e. do not generate SCI event) while the rest of handlers complete successfully mapping device in address space and whatever else.

This will have a well prepared device in system but guest is not notified, right?

yes, it's theoretically possible to move check up the call stack to machine level and not call secondary hotplug handler on non hotplugged device but that again depends on what secondary hotplug handler does. I'd rather keep logic independent here unless there is good reason not to do so.

Quoted 3 lines

My idea to move the check in pre-plug will result the qemu return when it see no SCI event will be generated, so there is no device created.

Quoted 7 lines
Quoted 1 line

I guess current behavior is a designed decision?

I'd say so.
PS:
QEMU's hotplug_hadler API is misnamed at this point as it handles both cold (basic device wiring) and hotplug (processing hotplug). Maybe we should rename it but I'm not irritated enough by it to do so.

After re-read the code, I found something more.

First let me describe my understanding a bit.

To hotplug a device, several part are related:

* device itself
* machine's hotplug_handler
* bus's hotplug_handler
* acpi configuration

Currently, some checks are mixed, which makes the logic not that clear.

Let's come back to the problem in this thread.

The check in concern is the device's hotpluggable property. And when we look into device_set_realized, this property is already checked at the very beginning. This means when we go all the way down to acpi_memory_plug_cb(), if this device is un-hotpluggable, it is must not a hotplugged device. And the acpi_send_event() will not be triggered.

Based on my understanding, mdev->dimm and mdev->is_enabld looks still necessary to be set for a un-hotplugged device. For both hotpluggable and un-hotpluggable device. Skip these two steps if the device is not hotpluggable looks not consistent with others?

Signature · 2 lines
Wei Yang
Help you, Help me
#19 lore raw
Quoted 73 lines · On Thu, 28 Feb 2019 08:46:10 +0800 Wei Yang <[email protected]>…
Quoted 43 lines · Igor Mammedov
Quoted 23 lines · On Wed, 27 Feb 2019 13:59:20 +0000 Wei Yang <[email protected]> wrote:
Quoted 22 lines · Igor Mammedov
Quoted 11 lines · On Mon, 25 Feb 2019 12:47:14 +0000 Wei Yang <[email protected]> wrote:
Quoted 7 lines
Quoted 1 line

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Ok, if we'd support un-hotpluggable mem device, we could retain this check. But this maybe not a proper place.

Quoted 6 lines

Based on my understanding, at this point, every thing has been done properly. If this check pass, we will send an acpi interrupt to notify the guest. In case this is an un-hotpluggable device, every thing looks good but no effect in guest. Because we skip the notification.

Maybe we need to move the check in pre-plug?

And what would happen then and afterwards?

Quoted 7 lines
Quoted 4 lines

Point is to make one of the handlers in chain to ignore plug event (i.e. do not generate SCI event) while the rest of handlers complete successfully mapping device in address space and whatever else.

This will have a well prepared device in system but guest is not notified, right?

yes, it's theoretically possible to move check up the call stack to machine level and not call secondary hotplug handler on non hotplugged device but that again depends on what secondary hotplug handler does. I'd rather keep logic independent here unless there is good reason not to do so.

Quoted 3 lines

My idea to move the check in pre-plug will result the qemu return when it see no SCI event will be generated, so there is no device created.

Quoted 7 lines
Quoted 1 line

I guess current behavior is a designed decision?

I'd say so.
PS:
QEMU's hotplug_hadler API is misnamed at this point as it handles both cold (basic device wiring) and hotplug (processing hotplug). Maybe we should rename it but I'm not irritated enough by it to do so.

After re-read the code, I found something more.

First let me describe my understanding a bit.

To hotplug a device, several part are related:

* device itself
* machine's hotplug_handler
* bus's hotplug_handler
* acpi configuration

Currently, some checks are mixed, which makes the logic not that clear.

Let's come back to the problem in this thread.

The check in concern is the device's hotpluggable property. And when we look into device_set_realized, this property is already checked at the very beginning. This means when we go all the way down to acpi_memory_plug_cb(), if this device is un-hotpluggable, it is must not a hotplugged device. And the acpi_send_event() will not be triggered.

Quoted 4 lines

Based on my understanding, mdev->dimm and mdev->is_enabld looks still necessary to be set for a un-hotplugged device. For both hotpluggable and un-hotpluggable device. Skip these two steps if the device is not hotpluggable looks not consistent with others?

it might be inconsistent and broken since we don't actually have a nonhotpluggable memory device yet. But I'd would fix it when such device is added (it might depend on being added device whether it needs to be tracked by acpi memory hotplug path or if it uses other means in which case check is correct)

author
#20 lore raw
Quoted 80 lines · Igor Mammedov
Quoted 73 lines · On Thu, 28 Feb 2019 08:46:10 +0800 Wei Yang <[email protected]>…
Quoted 43 lines · Igor Mammedov
Quoted 23 lines · On Wed, 27 Feb 2019 13:59:20 +0000 Wei Yang <[email protected]> wrote:
Quoted 22 lines · Igor Mammedov
Quoted 11 lines · On Mon, 25 Feb 2019 12:47:14 +0000 Wei Yang <[email protected]> wrote:
Quoted 7 lines
Quoted 1 line

To me, this is a general rule for PCDIMM, they are hotpluggable.

yes, PCDIMMs are hotpluggable but apci device (piix4pm/ich9/...) handling hotplug should ignore any non-hotpluggable one. If you remove check and then later someone else would add non-hotpluggable memory device or make PC-DIMMs or its variant of non-hotpluggable one, acpi device handling will break. Hence I'd prefer to keep the check.

Ok, if we'd support un-hotpluggable mem device, we could retain this check. But this maybe not a proper place.

Quoted 6 lines

Based on my understanding, at this point, every thing has been done properly. If this check pass, we will send an acpi interrupt to notify the guest. In case this is an un-hotpluggable device, every thing looks good but no effect in guest. Because we skip the notification.

Maybe we need to move the check in pre-plug?

And what would happen then and afterwards?

Quoted 7 lines
Quoted 4 lines

Point is to make one of the handlers in chain to ignore plug event (i.e. do not generate SCI event) while the rest of handlers complete successfully mapping device in address space and whatever else.

This will have a well prepared device in system but guest is not notified, right?

yes, it's theoretically possible to move check up the call stack to machine level and not call secondary hotplug handler on non hotplugged device but that again depends on what secondary hotplug handler does. I'd rather keep logic independent here unless there is good reason not to do so.

Quoted 3 lines

My idea to move the check in pre-plug will result the qemu return when it see no SCI event will be generated, so there is no device created.

Quoted 7 lines
Quoted 1 line

I guess current behavior is a designed decision?

I'd say so.
PS:
QEMU's hotplug_hadler API is misnamed at this point as it handles both cold (basic device wiring) and hotplug (processing hotplug). Maybe we should rename it but I'm not irritated enough by it to do so.

After re-read the code, I found something more.

First let me describe my understanding a bit.

To hotplug a device, several part are related:

* device itself
* machine's hotplug_handler
* bus's hotplug_handler
* acpi configuration

Currently, some checks are mixed, which makes the logic not that clear.

Let's come back to the problem in this thread.

The check in concern is the device's hotpluggable property. And when we look into device_set_realized, this property is already checked at the very beginning. This means when we go all the way down to acpi_memory_plug_cb(), if this device is un-hotpluggable, it is must not a hotplugged device. And the acpi_send_event() will not be triggered.

Quoted 4 lines

Based on my understanding, mdev->dimm and mdev->is_enabld looks still necessary to be set for a un-hotplugged device. For both hotpluggable and un-hotpluggable device. Skip these two steps if the device is not hotpluggable looks not consistent with others?

Quoted 6 lines

it might be inconsistent and broken since we don't actually have a nonhotpluggable memory device yet. But I'd would fix it when such device is added (it might depend on being added device whether it needs to be tracked by acpi memory hotplug path or if it uses other means in which case check is correct)

Ok, got your point.

Signature · 2 lines
Wei Yang
Help you, Help me
#21 lore raw
In reply to Wei Yang
Quoted 6 lines · Wei Yang

Three trivial cleanup for pc-dimm.

Patch [1] remove the check on class->hotpluggable since pc-dimm is always hotpluggable.
Patch [2] remove nvdimm_realize
Patch [2] remove pcdimm realize-callback

Please copy Igor on ACPI patches.

Thanks!

Quoted 17 lines

v2:
* fix warning in Patch 1
* split Patch 2 into two

Wei Yang (3):
pc-dimm: remove check on pc-dimm hotpluggable
mem/nvdimm: remove nvdimm_realize
pc-dimm: revert "introduce realize callback"

 hw/acpi/memory_hotplug.c |  5 -----
 hw/mem/nvdimm.c          | 11 -----------
 hw/mem/pc-dimm.c         |  5 -----
 include/hw/mem/pc-dimm.h |  3 ---
 4 files changed, 24 deletions(-)
Signature · 1 line
2.19.1