| Message ID | 20241025104103.342188-1-david@redhat.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [v1] virtio-mem: unplug memory only during system resets, not device resets | expand |
On 25.10.24 12:41, David Hildenbrand wrote: > We recently converted from the LegacyReset to the new reset framework > in commit c009a311e939 ("virtio-mem: Use new Resettable framework instead > of LegacyReset") to be able to use the ResetType to filter out wakeup > resets. > > However, this change had an undesired implications: as we override the > Resettable interface methods in VirtIOMEMClass, the reset handler will > not only get called during system resets (i.e., qemu_devices_reset()) > but also during any direct or indirect device rests (e.g., > device_cold_reset()). > > Further, we might now receive two reset callbacks during > qemu_devices_reset(), first when reset by a parent and later when reset > directly. > > The memory state of virtio-mem devices is rather special: it's supposed to > be persistent/unchanged during most resets (similar to resetting a hard > disk will not destroy the data), unless actually cold-resetting the whole > system (different to a hard disk where a reboot will not destroy the data): > ripping out system RAM is something guest OSes don't particularly enjoy, > but we want to detect when rebooting to an OS that does not support > virtio-mem and wouldn't be able to detect+use the memory -- and we want > to force-defragment hotplugged memory to also shrink the usable device > memory region. So we rally want to catch system resets to do that. > > On supported targets (e.g., x86), getting a cold reset on the > device/parent triggers is not that easy (but looks like PCI code > might trigger it), so this implication went unnoticed. > > However, with upcoming s390x support it is problematic: during > kdump, s390x triggers a subsystem reset, ending up in > s390_machine_reset() and calling only subsystem_reset() instead of > qemu_devices_reset() -- because it's not a full system reset. > > In subsystem_reset(), s390x performs a device_cold_reset() of any > TYPE_VIRTUAL_CSS_BRIDGE device, which ends up resetting all children, > including the virtio-mem device. Consequently, we wrongly detect a system > reset and unplug all device memory, resulting in hotplugged memory not > getting included in the crash dump -- undesired. > > We really must not mess with hotplugged memory state during simple > device resets. To fix, create+register a new reset object that will only > get triggered during qemu_devices_reset() calls, but not during any other > resets as it is logically not the child of any other object. > > Cc: "Michael S. Tsirkin" <mst@redhat.com> > Cc: Juraj Marcin <jmarcin@redhat.com> > Cc: Peter Maydell <peter.maydell@linaro.org> > Signed-off-by: David Hildenbrand <david@redhat.com> > --- If there are no further comments, I'll queue this as a preparation for s390x virtio-mem support.
On Fri, Oct 25, 2024 at 12:41:03PM +0200, David Hildenbrand wrote: > We recently converted from the LegacyReset to the new reset framework > in commit c009a311e939 ("virtio-mem: Use new Resettable framework instead > of LegacyReset") to be able to use the ResetType to filter out wakeup > resets. > > However, this change had an undesired implications: as we override the > Resettable interface methods in VirtIOMEMClass, the reset handler will > not only get called during system resets (i.e., qemu_devices_reset()) > but also during any direct or indirect device rests (e.g., > device_cold_reset()). > > Further, we might now receive two reset callbacks during > qemu_devices_reset(), first when reset by a parent and later when reset > directly. > > The memory state of virtio-mem devices is rather special: it's supposed to > be persistent/unchanged during most resets (similar to resetting a hard > disk will not destroy the data), unless actually cold-resetting the whole > system (different to a hard disk where a reboot will not destroy the data): > ripping out system RAM is something guest OSes don't particularly enjoy, > but we want to detect when rebooting to an OS that does not support > virtio-mem and wouldn't be able to detect+use the memory -- and we want > to force-defragment hotplugged memory to also shrink the usable device > memory region. So we rally want to catch system resets to do that. > > On supported targets (e.g., x86), getting a cold reset on the > device/parent triggers is not that easy (but looks like PCI code > might trigger it), so this implication went unnoticed. > > However, with upcoming s390x support it is problematic: during > kdump, s390x triggers a subsystem reset, ending up in > s390_machine_reset() and calling only subsystem_reset() instead of > qemu_devices_reset() -- because it's not a full system reset. > > In subsystem_reset(), s390x performs a device_cold_reset() of any > TYPE_VIRTUAL_CSS_BRIDGE device, which ends up resetting all children, > including the virtio-mem device. Consequently, we wrongly detect a system > reset and unplug all device memory, resulting in hotplugged memory not > getting included in the crash dump -- undesired. > > We really must not mess with hotplugged memory state during simple > device resets. To fix, create+register a new reset object that will only > get triggered during qemu_devices_reset() calls, but not during any other > resets as it is logically not the child of any other object. > > Cc: "Michael S. Tsirkin" <mst@redhat.com> > Cc: Juraj Marcin <jmarcin@redhat.com> > Cc: Peter Maydell <peter.maydell@linaro.org> > Signed-off-by: David Hildenbrand <david@redhat.com> Acked-by: Michael S. Tsirkin <mst@redhat.com> > --- > hw/virtio/virtio-mem.c | 103 +++++++++++++++++++++++---------- > include/hw/virtio/virtio-mem.h | 12 +++- > 2 files changed, 83 insertions(+), 32 deletions(-) > > diff --git a/hw/virtio/virtio-mem.c b/hw/virtio/virtio-mem.c > index ae1e81d7ba..08e0e9da1c 100644 > --- a/hw/virtio/virtio-mem.c > +++ b/hw/virtio/virtio-mem.c > @@ -956,6 +956,7 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp) > VirtIOMEM *vmem = VIRTIO_MEM(dev); > uint64_t page_size; > RAMBlock *rb; > + Object *obj; > int ret; > > if (!vmem->memdev) { > @@ -1121,7 +1122,28 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp) > vmstate_register_any(VMSTATE_IF(vmem), > &vmstate_virtio_mem_device_early, vmem); > } > - qemu_register_resettable(OBJECT(vmem)); > + > + /* > + * We only want to unplug all memory to start with a clean slate when > + * it is safe for the guest -- during system resets that call > + * qemu_devices_reset(). > + * > + * We'll filter out selected qemu_devices_reset() calls used for other > + * purposes, like resetting all devices during wakeup from suspend on > + * x86 based on the reset type passed to qemu_devices_reset(). > + * > + * Unplugging all memory during simple device resets can result in the VM > + * unexpectedly losing RAM, corrupting VM state. > + * > + * Simple device resets (or resets triggered by getting a parent device > + * reset) must not change the state of plugged memory blocks. Therefore, > + * we need a dedicated reset object that only gets called during > + * qemu_devices_reset(). > + */ > + obj = object_new(TYPE_VIRTIO_MEM_SYSTEM_RESET); > + vmem->system_reset = VIRTIO_MEM_SYSTEM_RESET(obj); > + vmem->system_reset->vmem = vmem; > + qemu_register_resettable(obj); > > /* > * Set ourselves as RamDiscardManager before the plug handler maps the > @@ -1141,7 +1163,10 @@ static void virtio_mem_device_unrealize(DeviceState *dev) > * found via an address space anymore. Unset ourselves. > */ > memory_region_set_ram_discard_manager(&vmem->memdev->mr, NULL); > - qemu_unregister_resettable(OBJECT(vmem)); > + > + qemu_unregister_resettable(OBJECT(vmem->system_reset)); > + object_unref(OBJECT(vmem->system_reset)); > + > if (vmem->early_migration) { > vmstate_unregister(VMSTATE_IF(vmem), &vmstate_virtio_mem_device_early, > vmem); > @@ -1841,38 +1866,12 @@ static void virtio_mem_unplug_request_check(VirtIOMEM *vmem, Error **errp) > } > } > > -static ResettableState *virtio_mem_get_reset_state(Object *obj) > -{ > - VirtIOMEM *vmem = VIRTIO_MEM(obj); > - return &vmem->reset_state; > -} > - > -static void virtio_mem_system_reset_hold(Object *obj, ResetType type) > -{ > - VirtIOMEM *vmem = VIRTIO_MEM(obj); > - > - /* > - * When waking up from standby/suspend-to-ram, do not unplug any memory. > - */ > - if (type == RESET_TYPE_WAKEUP) { > - return; > - } > - > - /* > - * During usual resets, we will unplug all memory and shrink the usable > - * region size. This is, however, not possible in all scenarios. Then, > - * the guest has to deal with this manually (VIRTIO_MEM_REQ_UNPLUG_ALL). > - */ > - virtio_mem_unplug_all(vmem); > -} > - > static void virtio_mem_class_init(ObjectClass *klass, void *data) > { > DeviceClass *dc = DEVICE_CLASS(klass); > VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass); > VirtIOMEMClass *vmc = VIRTIO_MEM_CLASS(klass); > RamDiscardManagerClass *rdmc = RAM_DISCARD_MANAGER_CLASS(klass); > - ResettableClass *rc = RESETTABLE_CLASS(klass); > > device_class_set_props(dc, virtio_mem_properties); > dc->vmsd = &vmstate_virtio_mem; > @@ -1899,9 +1898,6 @@ static void virtio_mem_class_init(ObjectClass *klass, void *data) > rdmc->replay_discarded = virtio_mem_rdm_replay_discarded; > rdmc->register_listener = virtio_mem_rdm_register_listener; > rdmc->unregister_listener = virtio_mem_rdm_unregister_listener; > - > - rc->get_state = virtio_mem_get_reset_state; > - rc->phases.hold = virtio_mem_system_reset_hold; > } > > static const TypeInfo virtio_mem_info = { > @@ -1924,3 +1920,48 @@ static void virtio_register_types(void) > } > > type_init(virtio_register_types) > + > +OBJECT_DEFINE_SIMPLE_TYPE_WITH_INTERFACES(VirtioMemSystemReset, virtio_mem_system_reset, VIRTIO_MEM_SYSTEM_RESET, OBJECT, { TYPE_RESETTABLE_INTERFACE }, { }) > + > +static void virtio_mem_system_reset_init(Object *obj) > +{ > +} > + > +static void virtio_mem_system_reset_finalize(Object *obj) > +{ > +} > + > +static ResettableState *virtio_mem_system_reset_get_state(Object *obj) > +{ > + VirtioMemSystemReset *vmem_reset = VIRTIO_MEM_SYSTEM_RESET(obj); > + > + return &vmem_reset->reset_state; > +} > + > +static void virtio_mem_system_reset_hold(Object *obj, ResetType type) > +{ > + VirtioMemSystemReset *vmem_reset = VIRTIO_MEM_SYSTEM_RESET(obj); > + VirtIOMEM *vmem = vmem_reset->vmem; > + > + /* > + * When waking up from standby/suspend-to-ram, do not unplug any memory. > + */ > + if (type == RESET_TYPE_WAKEUP) { > + return; > + } > + > + /* > + * During usual resets, we will unplug all memory and shrink the usable > + * region size. This is, however, not possible in all scenarios. Then, > + * the guest has to deal with this manually (VIRTIO_MEM_REQ_UNPLUG_ALL). > + */ > + virtio_mem_unplug_all(vmem); > +} > + > +static void virtio_mem_system_reset_class_init(ObjectClass *klass, void *data) > +{ > + ResettableClass *rc = RESETTABLE_CLASS(klass); > + > + rc->get_state = virtio_mem_system_reset_get_state; > + rc->phases.hold = virtio_mem_system_reset_hold; > +} > diff --git a/include/hw/virtio/virtio-mem.h b/include/hw/virtio/virtio-mem.h > index a1af144c28..abde1c4101 100644 > --- a/include/hw/virtio/virtio-mem.h > +++ b/include/hw/virtio/virtio-mem.h > @@ -25,6 +25,10 @@ > OBJECT_DECLARE_TYPE(VirtIOMEM, VirtIOMEMClass, > VIRTIO_MEM) > > +#define TYPE_VIRTIO_MEM_SYSTEM_RESET "virtio-mem-system-reset" > + > +OBJECT_DECLARE_SIMPLE_TYPE(VirtioMemSystemReset, VIRTIO_MEM_SYSTEM_RESET) > + > #define VIRTIO_MEM_MEMDEV_PROP "memdev" > #define VIRTIO_MEM_NODE_PROP "node" > #define VIRTIO_MEM_SIZE_PROP "size" > @@ -117,8 +121,14 @@ struct VirtIOMEM { > /* listeners to notify on plug/unplug activity. */ > QLIST_HEAD(, RamDiscardListener) rdl_list; > > - /* State of the resettable container */ > + /* Catch system resets -> qemu_devices_reset() only. */ > + VirtioMemSystemReset *system_reset; > +}; > + > +struct VirtioMemSystemReset { > + Object parent; > ResettableState reset_state; > + VirtIOMEM *vmem; > }; > > struct VirtIOMEMClass { > -- > 2.46.1
Hi David, On 25/10/24 11:41, David Hildenbrand wrote: > We recently converted from the LegacyReset to the new reset framework > in commit c009a311e939 ("virtio-mem: Use new Resettable framework instead > of LegacyReset") to be able to use the ResetType to filter out wakeup > resets. > > However, this change had an undesired implications: as we override the > Resettable interface methods in VirtIOMEMClass, the reset handler will > not only get called during system resets (i.e., qemu_devices_reset()) > but also during any direct or indirect device rests (e.g., > device_cold_reset()). > > Further, we might now receive two reset callbacks during > qemu_devices_reset(), first when reset by a parent and later when reset > directly. > > The memory state of virtio-mem devices is rather special: it's supposed to > be persistent/unchanged during most resets (similar to resetting a hard > disk will not destroy the data), unless actually cold-resetting the whole > system (different to a hard disk where a reboot will not destroy the data): > ripping out system RAM is something guest OSes don't particularly enjoy, > but we want to detect when rebooting to an OS that does not support > virtio-mem and wouldn't be able to detect+use the memory -- and we want > to force-defragment hotplugged memory to also shrink the usable device > memory region. So we rally want to catch system resets to do that. > > On supported targets (e.g., x86), getting a cold reset on the > device/parent triggers is not that easy (but looks like PCI code > might trigger it), so this implication went unnoticed. > > However, with upcoming s390x support it is problematic: during > kdump, s390x triggers a subsystem reset, ending up in > s390_machine_reset() and calling only subsystem_reset() instead of > qemu_devices_reset() -- because it's not a full system reset. > > In subsystem_reset(), s390x performs a device_cold_reset() of any > TYPE_VIRTUAL_CSS_BRIDGE device, which ends up resetting all children, > including the virtio-mem device. Consequently, we wrongly detect a system > reset and unplug all device memory, resulting in hotplugged memory not > getting included in the crash dump -- undesired. > > We really must not mess with hotplugged memory state during simple > device resets. To fix, create+register a new reset object that will only > get triggered during qemu_devices_reset() calls, but not during any other > resets as it is logically not the child of any other object. > > Cc: "Michael S. Tsirkin" <mst@redhat.com> > Cc: Juraj Marcin <jmarcin@redhat.com> > Cc: Peter Maydell <peter.maydell@linaro.org> > Signed-off-by: David Hildenbrand <david@redhat.com> > --- > hw/virtio/virtio-mem.c | 103 +++++++++++++++++++++++---------- > include/hw/virtio/virtio-mem.h | 12 +++- > 2 files changed, 83 insertions(+), 32 deletions(-) > +OBJECT_DEFINE_SIMPLE_TYPE_WITH_INTERFACES(VirtioMemSystemReset, virtio_mem_system_reset, VIRTIO_MEM_SYSTEM_RESET, OBJECT, { TYPE_RESETTABLE_INTERFACE }, { }) Please fix style when applying :) > diff --git a/include/hw/virtio/virtio-mem.h b/include/hw/virtio/virtio-mem.h > index a1af144c28..abde1c4101 100644 > --- a/include/hw/virtio/virtio-mem.h > +++ b/include/hw/virtio/virtio-mem.h > +struct VirtioMemSystemReset { > + Object parent; And add a newline here, thanks! > ResettableState reset_state; > + VirtIOMEM *vmem; > }; > > struct VirtIOMEMClass {
On 08.11.24 15:37, Philippe Mathieu-Daudé wrote: > Hi David, > > On 25/10/24 11:41, David Hildenbrand wrote: >> We recently converted from the LegacyReset to the new reset framework >> in commit c009a311e939 ("virtio-mem: Use new Resettable framework instead >> of LegacyReset") to be able to use the ResetType to filter out wakeup >> resets. >> >> However, this change had an undesired implications: as we override the >> Resettable interface methods in VirtIOMEMClass, the reset handler will >> not only get called during system resets (i.e., qemu_devices_reset()) >> but also during any direct or indirect device rests (e.g., >> device_cold_reset()). >> >> Further, we might now receive two reset callbacks during >> qemu_devices_reset(), first when reset by a parent and later when reset >> directly. >> >> The memory state of virtio-mem devices is rather special: it's supposed to >> be persistent/unchanged during most resets (similar to resetting a hard >> disk will not destroy the data), unless actually cold-resetting the whole >> system (different to a hard disk where a reboot will not destroy the data): >> ripping out system RAM is something guest OSes don't particularly enjoy, >> but we want to detect when rebooting to an OS that does not support >> virtio-mem and wouldn't be able to detect+use the memory -- and we want >> to force-defragment hotplugged memory to also shrink the usable device >> memory region. So we rally want to catch system resets to do that. >> >> On supported targets (e.g., x86), getting a cold reset on the >> device/parent triggers is not that easy (but looks like PCI code >> might trigger it), so this implication went unnoticed. >> >> However, with upcoming s390x support it is problematic: during >> kdump, s390x triggers a subsystem reset, ending up in >> s390_machine_reset() and calling only subsystem_reset() instead of >> qemu_devices_reset() -- because it's not a full system reset. >> >> In subsystem_reset(), s390x performs a device_cold_reset() of any >> TYPE_VIRTUAL_CSS_BRIDGE device, which ends up resetting all children, >> including the virtio-mem device. Consequently, we wrongly detect a system >> reset and unplug all device memory, resulting in hotplugged memory not >> getting included in the crash dump -- undesired. >> >> We really must not mess with hotplugged memory state during simple >> device resets. To fix, create+register a new reset object that will only >> get triggered during qemu_devices_reset() calls, but not during any other >> resets as it is logically not the child of any other object. >> >> Cc: "Michael S. Tsirkin" <mst@redhat.com> >> Cc: Juraj Marcin <jmarcin@redhat.com> >> Cc: Peter Maydell <peter.maydell@linaro.org> >> Signed-off-by: David Hildenbrand <david@redhat.com> >> --- >> hw/virtio/virtio-mem.c | 103 +++++++++++++++++++++++---------- >> include/hw/virtio/virtio-mem.h | 12 +++- >> 2 files changed, 83 insertions(+), 32 deletions(-) > > >> +OBJECT_DEFINE_SIMPLE_TYPE_WITH_INTERFACES(VirtioMemSystemReset, virtio_mem_system_reset, VIRTIO_MEM_SYSTEM_RESET, OBJECT, { TYPE_RESETTABLE_INTERFACE }, { }) > > Please fix style when applying :) The long line? I used hw/core/reset.c and hw/core/resetcontainer.c as inspiration -- they surely can't be wrong ;) > >> diff --git a/include/hw/virtio/virtio-mem.h b/include/hw/virtio/virtio-mem.h >> index a1af144c28..abde1c4101 100644 >> --- a/include/hw/virtio/virtio-mem.h >> +++ b/include/hw/virtio/virtio-mem.h > > >> +struct VirtioMemSystemReset { >> + Object parent; > > And add a newline here, thanks! Make sense, thanks!
diff --git a/hw/virtio/virtio-mem.c b/hw/virtio/virtio-mem.c index ae1e81d7ba..08e0e9da1c 100644 --- a/hw/virtio/virtio-mem.c +++ b/hw/virtio/virtio-mem.c @@ -956,6 +956,7 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp) VirtIOMEM *vmem = VIRTIO_MEM(dev); uint64_t page_size; RAMBlock *rb; + Object *obj; int ret; if (!vmem->memdev) { @@ -1121,7 +1122,28 @@ static void virtio_mem_device_realize(DeviceState *dev, Error **errp) vmstate_register_any(VMSTATE_IF(vmem), &vmstate_virtio_mem_device_early, vmem); } - qemu_register_resettable(OBJECT(vmem)); + + /* + * We only want to unplug all memory to start with a clean slate when + * it is safe for the guest -- during system resets that call + * qemu_devices_reset(). + * + * We'll filter out selected qemu_devices_reset() calls used for other + * purposes, like resetting all devices during wakeup from suspend on + * x86 based on the reset type passed to qemu_devices_reset(). + * + * Unplugging all memory during simple device resets can result in the VM + * unexpectedly losing RAM, corrupting VM state. + * + * Simple device resets (or resets triggered by getting a parent device + * reset) must not change the state of plugged memory blocks. Therefore, + * we need a dedicated reset object that only gets called during + * qemu_devices_reset(). + */ + obj = object_new(TYPE_VIRTIO_MEM_SYSTEM_RESET); + vmem->system_reset = VIRTIO_MEM_SYSTEM_RESET(obj); + vmem->system_reset->vmem = vmem; + qemu_register_resettable(obj); /* * Set ourselves as RamDiscardManager before the plug handler maps the @@ -1141,7 +1163,10 @@ static void virtio_mem_device_unrealize(DeviceState *dev) * found via an address space anymore. Unset ourselves. */ memory_region_set_ram_discard_manager(&vmem->memdev->mr, NULL); - qemu_unregister_resettable(OBJECT(vmem)); + + qemu_unregister_resettable(OBJECT(vmem->system_reset)); + object_unref(OBJECT(vmem->system_reset)); + if (vmem->early_migration) { vmstate_unregister(VMSTATE_IF(vmem), &vmstate_virtio_mem_device_early, vmem); @@ -1841,38 +1866,12 @@ static void virtio_mem_unplug_request_check(VirtIOMEM *vmem, Error **errp) } } -static ResettableState *virtio_mem_get_reset_state(Object *obj) -{ - VirtIOMEM *vmem = VIRTIO_MEM(obj); - return &vmem->reset_state; -} - -static void virtio_mem_system_reset_hold(Object *obj, ResetType type) -{ - VirtIOMEM *vmem = VIRTIO_MEM(obj); - - /* - * When waking up from standby/suspend-to-ram, do not unplug any memory. - */ - if (type == RESET_TYPE_WAKEUP) { - return; - } - - /* - * During usual resets, we will unplug all memory and shrink the usable - * region size. This is, however, not possible in all scenarios. Then, - * the guest has to deal with this manually (VIRTIO_MEM_REQ_UNPLUG_ALL). - */ - virtio_mem_unplug_all(vmem); -} - static void virtio_mem_class_init(ObjectClass *klass, void *data) { DeviceClass *dc = DEVICE_CLASS(klass); VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass); VirtIOMEMClass *vmc = VIRTIO_MEM_CLASS(klass); RamDiscardManagerClass *rdmc = RAM_DISCARD_MANAGER_CLASS(klass); - ResettableClass *rc = RESETTABLE_CLASS(klass); device_class_set_props(dc, virtio_mem_properties); dc->vmsd = &vmstate_virtio_mem; @@ -1899,9 +1898,6 @@ static void virtio_mem_class_init(ObjectClass *klass, void *data) rdmc->replay_discarded = virtio_mem_rdm_replay_discarded; rdmc->register_listener = virtio_mem_rdm_register_listener; rdmc->unregister_listener = virtio_mem_rdm_unregister_listener; - - rc->get_state = virtio_mem_get_reset_state; - rc->phases.hold = virtio_mem_system_reset_hold; } static const TypeInfo virtio_mem_info = { @@ -1924,3 +1920,48 @@ static void virtio_register_types(void) } type_init(virtio_register_types) + +OBJECT_DEFINE_SIMPLE_TYPE_WITH_INTERFACES(VirtioMemSystemReset, virtio_mem_system_reset, VIRTIO_MEM_SYSTEM_RESET, OBJECT, { TYPE_RESETTABLE_INTERFACE }, { }) + +static void virtio_mem_system_reset_init(Object *obj) +{ +} + +static void virtio_mem_system_reset_finalize(Object *obj) +{ +} + +static ResettableState *virtio_mem_system_reset_get_state(Object *obj) +{ + VirtioMemSystemReset *vmem_reset = VIRTIO_MEM_SYSTEM_RESET(obj); + + return &vmem_reset->reset_state; +} + +static void virtio_mem_system_reset_hold(Object *obj, ResetType type) +{ + VirtioMemSystemReset *vmem_reset = VIRTIO_MEM_SYSTEM_RESET(obj); + VirtIOMEM *vmem = vmem_reset->vmem; + + /* + * When waking up from standby/suspend-to-ram, do not unplug any memory. + */ + if (type == RESET_TYPE_WAKEUP) { + return; + } + + /* + * During usual resets, we will unplug all memory and shrink the usable + * region size. This is, however, not possible in all scenarios. Then, + * the guest has to deal with this manually (VIRTIO_MEM_REQ_UNPLUG_ALL). + */ + virtio_mem_unplug_all(vmem); +} + +static void virtio_mem_system_reset_class_init(ObjectClass *klass, void *data) +{ + ResettableClass *rc = RESETTABLE_CLASS(klass); + + rc->get_state = virtio_mem_system_reset_get_state; + rc->phases.hold = virtio_mem_system_reset_hold; +} diff --git a/include/hw/virtio/virtio-mem.h b/include/hw/virtio/virtio-mem.h index a1af144c28..abde1c4101 100644 --- a/include/hw/virtio/virtio-mem.h +++ b/include/hw/virtio/virtio-mem.h @@ -25,6 +25,10 @@ OBJECT_DECLARE_TYPE(VirtIOMEM, VirtIOMEMClass, VIRTIO_MEM) +#define TYPE_VIRTIO_MEM_SYSTEM_RESET "virtio-mem-system-reset" + +OBJECT_DECLARE_SIMPLE_TYPE(VirtioMemSystemReset, VIRTIO_MEM_SYSTEM_RESET) + #define VIRTIO_MEM_MEMDEV_PROP "memdev" #define VIRTIO_MEM_NODE_PROP "node" #define VIRTIO_MEM_SIZE_PROP "size" @@ -117,8 +121,14 @@ struct VirtIOMEM { /* listeners to notify on plug/unplug activity. */ QLIST_HEAD(, RamDiscardListener) rdl_list; - /* State of the resettable container */ + /* Catch system resets -> qemu_devices_reset() only. */ + VirtioMemSystemReset *system_reset; +}; + +struct VirtioMemSystemReset { + Object parent; ResettableState reset_state; + VirtIOMEM *vmem; }; struct VirtIOMEMClass {
We recently converted from the LegacyReset to the new reset framework in commit c009a311e939 ("virtio-mem: Use new Resettable framework instead of LegacyReset") to be able to use the ResetType to filter out wakeup resets. However, this change had an undesired implications: as we override the Resettable interface methods in VirtIOMEMClass, the reset handler will not only get called during system resets (i.e., qemu_devices_reset()) but also during any direct or indirect device rests (e.g., device_cold_reset()). Further, we might now receive two reset callbacks during qemu_devices_reset(), first when reset by a parent and later when reset directly. The memory state of virtio-mem devices is rather special: it's supposed to be persistent/unchanged during most resets (similar to resetting a hard disk will not destroy the data), unless actually cold-resetting the whole system (different to a hard disk where a reboot will not destroy the data): ripping out system RAM is something guest OSes don't particularly enjoy, but we want to detect when rebooting to an OS that does not support virtio-mem and wouldn't be able to detect+use the memory -- and we want to force-defragment hotplugged memory to also shrink the usable device memory region. So we rally want to catch system resets to do that. On supported targets (e.g., x86), getting a cold reset on the device/parent triggers is not that easy (but looks like PCI code might trigger it), so this implication went unnoticed. However, with upcoming s390x support it is problematic: during kdump, s390x triggers a subsystem reset, ending up in s390_machine_reset() and calling only subsystem_reset() instead of qemu_devices_reset() -- because it's not a full system reset. In subsystem_reset(), s390x performs a device_cold_reset() of any TYPE_VIRTUAL_CSS_BRIDGE device, which ends up resetting all children, including the virtio-mem device. Consequently, we wrongly detect a system reset and unplug all device memory, resulting in hotplugged memory not getting included in the crash dump -- undesired. We really must not mess with hotplugged memory state during simple device resets. To fix, create+register a new reset object that will only get triggered during qemu_devices_reset() calls, but not during any other resets as it is logically not the child of any other object. Cc: "Michael S. Tsirkin" <mst@redhat.com> Cc: Juraj Marcin <jmarcin@redhat.com> Cc: Peter Maydell <peter.maydell@linaro.org> Signed-off-by: David Hildenbrand <david@redhat.com> --- hw/virtio/virtio-mem.c | 103 +++++++++++++++++++++++---------- include/hw/virtio/virtio-mem.h | 12 +++- 2 files changed, 83 insertions(+), 32 deletions(-)