| Message ID | 20220225124848.909093-1-Jason@zx2c4.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [v4] virt: vmgenid: introduce driver for reinitializing RNG on VM fork | expand |
On Fri, Feb 25, 2022 at 01:48:48PM +0100, Jason A. Donenfeld wrote: > VM Generation ID is a feature from Microsoft, described at > <https://go.microsoft.com/fwlink/?LinkId=260709>, and supported by > Hyper-V and QEMU. Its usage is described in Microsoft's RNG whitepaper, > <https://aka.ms/win10rng>, as: > > If the OS is running in a VM, there is a problem that most > hypervisors can snapshot the state of the machine and later rewind > the VM state to the saved state. This results in the machine running > a second time with the exact same RNG state, which leads to serious > security problems. To reduce the window of vulnerability, Windows > 10 on a Hyper-V VM will detect when the VM state is reset, retrieve > a unique (not random) value from the hypervisor, and reseed the root > RNG with that unique value. This does not eliminate the > vulnerability, but it greatly reduces the time during which the RNG > system will produce the same outputs as it did during a previous > instantiation of the same VM state. > > Linux has the same issue, and given that vmgenid is supported already by > multiple hypervisors, we can implement more or less the same solution. > So this commit wires up the vmgenid ACPI notification to the RNG's newly > added add_vmfork_randomness() function. > > It can be used from qemu via the `-device vmgenid,guid=auto` parameter. > After setting that, use `savevm` in the monitor to save the VM state, > then quit QEMU, start it again, and use `loadvm`. That will trigger this > driver's notify function, which hands the new UUID to the RNG. This is > described in <https://git.qemu.org/?p=qemu.git;a=blob;f=docs/specs/vmgenid.txt>. > And there are hooks for this in libvirt as well, described in > <https://libvirt.org/formatdomain.html#general-metadata>. > > Note, however, that the treatment of this as a UUID is considered to be > an accidental QEMU nuance, per > <https://github.com/libguestfs/virt-v2v/blob/master/docs/vm-generation-id-across-hypervisors.txt>, > so this driver simply treats these bytes as an opaque 128-bit binary > blob, as per the spec. This doesn't really make a difference anyway, > considering that's how it ends up when handed to the RNG in the end. > > Cc: Adrian Catangiu <adrian@parity.io> > Cc: Daniel P. Berrangé <berrange@redhat.com> > Cc: Dominik Brodowski <linux@dominikbrodowski.net> > Cc: Ard Biesheuvel <ardb@kernel.org> > Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> > Reviewed-by: Laszlo Ersek <lersek@redhat.com> > Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
On Fri, Feb 25, 2022 at 01:48:48PM +0100, Jason A. Donenfeld wrote: > +static struct acpi_driver acpi_driver = { > + .name = "vmgenid", > + .ids = vmgenid_ids, > + .owner = THIS_MODULE, > + .ops = { > + .add = vmgenid_acpi_add, > + .notify = vmgenid_acpi_notify, > + } > +}; > + > +static int __init vmgenid_init(void) > +{ > + return acpi_bus_register_driver(&acpi_driver); > +} > + > +static void __exit vmgenid_exit(void) > +{ > + acpi_bus_unregister_driver(&acpi_driver); > +} > + > +module_init(vmgenid_init); > +module_exit(vmgenid_exit); Nit, you could use module_acpi_driver() to make this even smaller if you want to. thanks, greg k-h
On Fri, Feb 25, 2022 at 1:53 PM Greg KH <gregkh@linuxfoundation.org> wrote: > > On Fri, Feb 25, 2022 at 01:48:48PM +0100, Jason A. Donenfeld wrote: > > +static struct acpi_driver acpi_driver = { > > + .name = "vmgenid", > > + .ids = vmgenid_ids, > > + .owner = THIS_MODULE, > > + .ops = { > > + .add = vmgenid_acpi_add, > > + .notify = vmgenid_acpi_notify, > > + } > > +}; > > + > > +static int __init vmgenid_init(void) > > +{ > > + return acpi_bus_register_driver(&acpi_driver); > > +} > > + > > +static void __exit vmgenid_exit(void) > > +{ > > + acpi_bus_unregister_driver(&acpi_driver); > > +} > > + > > +module_init(vmgenid_init); > > +module_exit(vmgenid_exit); > > Nit, you could use module_acpi_driver() to make this even smaller if you > want to. Nice! Will do. Jason
On 25.02.22 13:48, Jason A. Donenfeld wrote: > > VM Generation ID is a feature from Microsoft, described at > <https://go.microsoft.com/fwlink/?LinkId=260709>, and supported by > Hyper-V and QEMU. Its usage is described in Microsoft's RNG whitepaper, > <https://aka.ms/win10rng>, as: > > If the OS is running in a VM, there is a problem that most > hypervisors can snapshot the state of the machine and later rewind > the VM state to the saved state. This results in the machine running > a second time with the exact same RNG state, which leads to serious > security problems. To reduce the window of vulnerability, Windows > 10 on a Hyper-V VM will detect when the VM state is reset, retrieve > a unique (not random) value from the hypervisor, and reseed the root > RNG with that unique value. This does not eliminate the > vulnerability, but it greatly reduces the time during which the RNG > system will produce the same outputs as it did during a previous > instantiation of the same VM state. > > Linux has the same issue, and given that vmgenid is supported already by > multiple hypervisors, we can implement more or less the same solution. > So this commit wires up the vmgenid ACPI notification to the RNG's newly > added add_vmfork_randomness() function. > > It can be used from qemu via the `-device vmgenid,guid=auto` parameter. > After setting that, use `savevm` in the monitor to save the VM state, > then quit QEMU, start it again, and use `loadvm`. That will trigger this > driver's notify function, which hands the new UUID to the RNG. This is > described in <https://git.qemu.org/?p=qemu.git;a=blob;f=docs/specs/vmgenid.txt>. > And there are hooks for this in libvirt as well, described in > <https://libvirt.org/formatdomain.html#general-metadata>. > > Note, however, that the treatment of this as a UUID is considered to be > an accidental QEMU nuance, per > <https://github.com/libguestfs/virt-v2v/blob/master/docs/vm-generation-id-across-hypervisors.txt>, > so this driver simply treats these bytes as an opaque 128-bit binary > blob, as per the spec. This doesn't really make a difference anyway, > considering that's how it ends up when handed to the RNG in the end. > > Cc: Adrian Catangiu <adrian@parity.io> > Cc: Daniel P. Berrangé <berrange@redhat.com> > Cc: Dominik Brodowski <linux@dominikbrodowski.net> > Cc: Ard Biesheuvel <ardb@kernel.org> > Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> > Reviewed-by: Laszlo Ersek <lersek@redhat.com> > Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> > --- > Changes v3->v4: > - Add this driver to MAINTAINERS, per Ard's request. > Note: I didn't really want to do this at first, because I was hoping the > original Amazon team looking into this last year would step up. But it seems > like that team has moved on, and anyway I've basically rewritten the driver > from scratch at this point -- not a single line of the original exists -- > and so I guess I'll maintain it myself. Adding Greg to the CC for his ack on > this. > - Don't use a static global state in case there are multiple instances. > - Use devm_memremap instead of the acpi internal functions. > - Default to being modular instead of a built-in, as apparently this is > udev-able. > > MAINTAINERS | 1 + > drivers/virt/Kconfig | 9 ++++ > drivers/virt/Makefile | 1 + > drivers/virt/vmgenid.c | 112 +++++++++++++++++++++++++++++++++++++++++ > 4 files changed, 123 insertions(+) > create mode 100644 drivers/virt/vmgenid.c > > diff --git a/MAINTAINERS b/MAINTAINERS > index 777cd6fa2b3d..a10997e15146 100644 > --- a/MAINTAINERS > +++ b/MAINTAINERS > @@ -16211,6 +16211,7 @@ M: Jason A. Donenfeld <Jason@zx2c4.com> > T: git https://git.kernel.org/pub/scm/linux/kernel/git/crng/random.git > S: Maintained > F: drivers/char/random.c > +F: drivers/virt/vmgenid.c > > RAPIDIO SUBSYSTEM > M: Matt Porter <mporter@kernel.crashing.org> > diff --git a/drivers/virt/Kconfig b/drivers/virt/Kconfig > index 8061e8ef449f..5596c7313f59 100644 > --- a/drivers/virt/Kconfig > +++ b/drivers/virt/Kconfig > @@ -13,6 +13,15 @@ menuconfig VIRT_DRIVERS > > if VIRT_DRIVERS > > +config VMGENID > + tristate "Virtual Machine Generation ID driver" > + default m > + depends on ACPI > + help > + Say Y here to use the hypervisor-provided Virtual Machine Generation ID > + to reseed the RNG when the VM is cloned. This is highly recommended if > + you intend to do any rollback / cloning / snapshotting of VMs. > + > config FSL_HV_MANAGER > tristate "Freescale hypervisor management driver" > depends on FSL_SOC > diff --git a/drivers/virt/Makefile b/drivers/virt/Makefile > index 3e272ea60cd9..108d0ffcc9aa 100644 > --- a/drivers/virt/Makefile > +++ b/drivers/virt/Makefile > @@ -4,6 +4,7 @@ > # > > obj-$(CONFIG_FSL_HV_MANAGER) += fsl_hypervisor.o > +obj-$(CONFIG_VMGENID) += vmgenid.o > obj-y += vboxguest/ > > obj-$(CONFIG_NITRO_ENCLAVES) += nitro_enclaves/ > diff --git a/drivers/virt/vmgenid.c b/drivers/virt/vmgenid.c > new file mode 100644 > index 000000000000..e3dd4afb33c6 > --- /dev/null > +++ b/drivers/virt/vmgenid.c > @@ -0,0 +1,112 @@ > +// SPDX-License-Identifier: GPL-2.0 > +/* > + * Copyright (C) 2022 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved. > + * > + * The "Virtual Machine Generation ID" is exposed via ACPI and changes when a > + * virtual machine forks or is cloned. This driver exists for shepherding that > + * information to random.c. > + */ > + > +#include <linux/kernel.h> > +#include <linux/module.h> > +#include <linux/acpi.h> > +#include <linux/random.h> > + > +ACPI_MODULE_NAME("vmgenid"); > + > +enum { VMGENID_SIZE = 16 }; > + > +struct vmgenid_state { > + u8 *next_id; > + u8 this_id[VMGENID_SIZE]; > +}; > + > +static int vmgenid_acpi_add(struct acpi_device *device) > +{ > + struct acpi_buffer parsed = { ACPI_ALLOCATE_BUFFER }; > + struct vmgenid_state *state; > + union acpi_object *obj; > + phys_addr_t phys_addr; > + acpi_status status; > + int ret = 0; > + > + state = devm_kmalloc(&device->dev, sizeof(*state), GFP_KERNEL); > + if (!state) > + return -ENOMEM; > + > + status = acpi_evaluate_object(device->handle, "ADDR", NULL, &parsed); > + if (ACPI_FAILURE(status)) { > + ACPI_EXCEPTION((AE_INFO, status, "Evaluating ADDR")); > + return -ENODEV; > + } > + obj = parsed.pointer; > + if (!obj || obj->type != ACPI_TYPE_PACKAGE || obj->package.count != 2 || > + obj->package.elements[0].type != ACPI_TYPE_INTEGER || > + obj->package.elements[1].type != ACPI_TYPE_INTEGER) { > + ret = -EINVAL; > + goto out; > + } > + > + phys_addr = (obj->package.elements[0].integer.value << 0) | > + (obj->package.elements[1].integer.value << 32); > + state->next_id = devm_memremap(&device->dev, phys_addr, VMGENID_SIZE, MEMREMAP_WB); > + if (!state->next_id) { > + ret = -ENOMEM; > + goto out; > + } > + > + memcpy(state->this_id, state->next_id, sizeof(state->this_id)); > + add_device_randomness(state->this_id, sizeof(state->this_id)); Please expose the vmgenid via /sysfs so that user space even remotely has a chance to check if it's been cloned. > + > + device->driver_data = state; > + > +out: > + ACPI_FREE(parsed.pointer); > + return ret; > +} > + > +static void vmgenid_acpi_notify(struct acpi_device *device, u32 event) > +{ > + struct vmgenid_state *state = acpi_driver_data(device); > + u8 old_id[VMGENID_SIZE]; > + > + memcpy(old_id, state->this_id, sizeof(old_id)); > + memcpy(state->this_id, state->next_id, sizeof(state->this_id)); > + if (!memcmp(old_id, state->this_id, sizeof(old_id))) > + return; > + add_vmfork_randomness(state->this_id, sizeof(state->this_id)); > +} > + > +static const struct acpi_device_id vmgenid_ids[] = { > + { "VMGENID", 0 }, > + { "QEMUVGID", 0 }, According to the VMGenID spec[1], you can only rely on _CID and _DDN for matching. They both contain "VM_Gen_Counter". The list above contains _HID values which are not an official identifier for the VMGenID device. IIRC the ACPI device match logic does match _CID in addition to _HID. However, it is limited to 8 characters. Let me paste an experimental hack I did back then to do the _CID matching instead. [1] https://download.microsoft.com/download/3/1/C/31CFC307-98CA-4CA5-914C-D9772691E214/VirtualMachineGenerationID.docx Alex diff --git a/drivers/acpi/bus.c b/drivers/acpi/bus.c index 1682f8b454a2..452443d79d87 100644 --- a/drivers/acpi/bus.c +++ b/drivers/acpi/bus.c @@ -748,7 +748,7 @@ static bool __acpi_match_device(struct acpi_device *device, /* First, check the ACPI/PNP IDs provided by the caller. */ if (acpi_ids) { for (id = acpi_ids; id->id[0] || id->cls; id++) { - if (id->id[0] && !strcmp((char *)id->id, hwid->id)) + if (id->id[0] && !strncmp((char *)id->id, hwid->id, ACPI_ID_LEN - 1)) goto out_acpi_match; if (id->cls && __acpi_match_device_cls(id, hwid)) goto out_acpi_match; diff --git a/drivers/virt/vmgenid.c b/drivers/virt/vmgenid.c index 75a787da8aad..0bfa422cf094 100644 --- a/drivers/virt/vmgenid.c +++ b/drivers/virt/vmgenid.c @@ -356,7 +356,8 @@ static void vmgenid_acpi_notify(struct acpi_device *device, u32 event) } static const struct acpi_device_id vmgenid_ids[] = { - {"QEMUVGID", 0}, + /* This really is VM_Gen_Counter, but we can only match 8 characters */ + {"VM_GEN_C", 0}, {"", 0}, }; Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
Hi Alex, On Fri, Feb 25, 2022 at 02:57:38PM +0100, Alexander Graf wrote: > > +static const struct acpi_device_id vmgenid_ids[] = { > > + { "VMGENID", 0 }, > > + { "QEMUVGID", 0 }, > > > According to the VMGenID spec[1], you can only rely on _CID and _DDN for > matching. They both contain "VM_Gen_Counter". The list above contains > _HID values which are not an official identifier for the VMGenID device. > > IIRC the ACPI device match logic does match _CID in addition to _HID. > However, it is limited to 8 characters. Let me paste an experimental > hack I did back then to do the _CID matching instead. > > [1] > https://download.microsoft.com/download/3/1/C/31CFC307-98CA-4CA5-914C-D9772691E214/VirtualMachineGenerationID.docx > > > Alex > > diff --git a/drivers/acpi/bus.c b/drivers/acpi/bus.c > index 1682f8b454a2..452443d79d87 100644 > --- a/drivers/acpi/bus.c > +++ b/drivers/acpi/bus.c > @@ -748,7 +748,7 @@ static bool __acpi_match_device(struct acpi_device > *device, > /* First, check the ACPI/PNP IDs provided by the caller. */ > if (acpi_ids) { > for (id = acpi_ids; id->id[0] || id->cls; id++) { > - if (id->id[0] && !strcmp((char *)id->id, hwid->id)) > + if (id->id[0] && !strncmp((char *)id->id, hwid->id, > ACPI_ID_LEN - 1)) > goto out_acpi_match; > if (id->cls && __acpi_match_device_cls(id, hwid)) > goto out_acpi_match; > diff --git a/drivers/virt/vmgenid.c b/drivers/virt/vmgenid.c > index 75a787da8aad..0bfa422cf094 100644 > --- a/drivers/virt/vmgenid.c > +++ b/drivers/virt/vmgenid.c > @@ -356,7 +356,8 @@ static void vmgenid_acpi_notify(struct acpi_device > *device, u32 event) > } > > static const struct acpi_device_id vmgenid_ids[] = { > - {"QEMUVGID", 0}, > + /* This really is VM_Gen_Counter, but we can only match 8 characters */ > + {"VM_GEN_C", 0}, > {"", 0}, > }; I recall this part of the old thread. From what I understood, using "VMGENID" + "QEMUVGID" worked /well enough/, even if that wasn't technically in-spec. Ard noted that relying on _CID like that is technically an ACPI spec notification. So we're between one spec and another, basically, and doing "VMGENID" + "QEMUVGID" requires fewer changes, as mentioned, appears to work fine in my testing. However, with that said, I think supporting this via "VM_Gen_Counter" would be a better eventual thing to do, but will require acks and changes from the ACPI maintainers. Do you think you could prepare your patch proposal above as something on-top of my tree [1]? And if you can convince the ACPI maintainers that that's okay, then I'll happily take the patch. Jason [1] https://git.kernel.org/pub/scm/linux/kernel/git/crng/random.git/
On Fri, Feb 25, 2022 at 3:12 PM Jason A. Donenfeld <Jason@zx2c4.com> wrote: > > Hi Alex, > > On Fri, Feb 25, 2022 at 02:57:38PM +0100, Alexander Graf wrote: > > > +static const struct acpi_device_id vmgenid_ids[] = { > > > + { "VMGENID", 0 }, > > > + { "QEMUVGID", 0 }, > > > > > > According to the VMGenID spec[1], you can only rely on _CID and _DDN for > > matching. They both contain "VM_Gen_Counter". The list above contains > > _HID values which are not an official identifier for the VMGenID device. > > > > IIRC the ACPI device match logic does match _CID in addition to _HID. > > However, it is limited to 8 characters. Let me paste an experimental > > hack I did back then to do the _CID matching instead. > > > > [1] > > https://download.microsoft.com/download/3/1/C/31CFC307-98CA-4CA5-914C-D9772691E214/VirtualMachineGenerationID.docx > > > > > > Alex > > > > diff --git a/drivers/acpi/bus.c b/drivers/acpi/bus.c > > index 1682f8b454a2..452443d79d87 100644 > > --- a/drivers/acpi/bus.c > > +++ b/drivers/acpi/bus.c > > @@ -748,7 +748,7 @@ static bool __acpi_match_device(struct acpi_device > > *device, > > /* First, check the ACPI/PNP IDs provided by the caller. */ > > if (acpi_ids) { > > for (id = acpi_ids; id->id[0] || id->cls; id++) { > > - if (id->id[0] && !strcmp((char *)id->id, hwid->id)) > > + if (id->id[0] && !strncmp((char *)id->id, hwid->id, > > ACPI_ID_LEN - 1)) > > goto out_acpi_match; > > if (id->cls && __acpi_match_device_cls(id, hwid)) > > goto out_acpi_match; > > diff --git a/drivers/virt/vmgenid.c b/drivers/virt/vmgenid.c > > index 75a787da8aad..0bfa422cf094 100644 > > --- a/drivers/virt/vmgenid.c > > +++ b/drivers/virt/vmgenid.c > > @@ -356,7 +356,8 @@ static void vmgenid_acpi_notify(struct acpi_device > > *device, u32 event) > > } > > > > static const struct acpi_device_id vmgenid_ids[] = { > > - {"QEMUVGID", 0}, > > + /* This really is VM_Gen_Counter, but we can only match 8 characters */ > > + {"VM_GEN_C", 0}, > > {"", 0}, > > }; > > I recall this part of the old thread. From what I understood, using > "VMGENID" + "QEMUVGID" worked /well enough/, even if that wasn't > technically in-spec. Ard noted that relying on _CID like that is > technically an ACPI spec notification. So we're between one spec and > another, basically, and doing "VMGENID" + "QEMUVGID" requires fewer > changes, as mentioned, appears to work fine in my testing. > > However, with that said, I think supporting this via "VM_Gen_Counter" > would be a better eventual thing to do, but will require acks and > changes from the ACPI maintainers. Do you think you could prepare your > patch proposal above as something on-top of my tree [1]? And if you can > convince the ACPI maintainers that that's okay, then I'll happily take > the patch. Closely related concern that whatever patch you come up with will have to handle is MODULE_DEVICE_TABLE and udev autoloading. I don't know if _CID matching is something that happens in udev or what its limits are, so that'll have to be researched and tested a bit. Jason
On 25.02.22 15:12, Jason A. Donenfeld wrote: > Hi Alex, > > On Fri, Feb 25, 2022 at 02:57:38PM +0100, Alexander Graf wrote: >>> +static const struct acpi_device_id vmgenid_ids[] = { >>> + { "VMGENID", 0 }, >>> + { "QEMUVGID", 0 }, >> >> According to the VMGenID spec[1], you can only rely on _CID and _DDN for >> matching. They both contain "VM_Gen_Counter". The list above contains >> _HID values which are not an official identifier for the VMGenID device. >> >> IIRC the ACPI device match logic does match _CID in addition to _HID. >> However, it is limited to 8 characters. Let me paste an experimental >> hack I did back then to do the _CID matching instead. >> >> [1] >> https://download.microsoft.com/download/3/1/C/31CFC307-98CA-4CA5-914C-D9772691E214/VirtualMachineGenerationID.docx >> >> >> Alex >> >> diff --git a/drivers/acpi/bus.c b/drivers/acpi/bus.c >> index 1682f8b454a2..452443d79d87 100644 >> --- a/drivers/acpi/bus.c >> +++ b/drivers/acpi/bus.c >> @@ -748,7 +748,7 @@ static bool __acpi_match_device(struct acpi_device >> *device, >> /* First, check the ACPI/PNP IDs provided by the caller. */ >> if (acpi_ids) { >> for (id = acpi_ids; id->id[0] || id->cls; id++) { >> - if (id->id[0] && !strcmp((char *)id->id, hwid->id)) >> + if (id->id[0] && !strncmp((char *)id->id, hwid->id, >> ACPI_ID_LEN - 1)) >> goto out_acpi_match; >> if (id->cls && __acpi_match_device_cls(id, hwid)) >> goto out_acpi_match; >> diff --git a/drivers/virt/vmgenid.c b/drivers/virt/vmgenid.c >> index 75a787da8aad..0bfa422cf094 100644 >> --- a/drivers/virt/vmgenid.c >> +++ b/drivers/virt/vmgenid.c >> @@ -356,7 +356,8 @@ static void vmgenid_acpi_notify(struct acpi_device >> *device, u32 event) >> } >> >> static const struct acpi_device_id vmgenid_ids[] = { >> - {"QEMUVGID", 0}, >> + /* This really is VM_Gen_Counter, but we can only match 8 characters */ >> + {"VM_GEN_C", 0}, >> {"", 0}, >> }; > I recall this part of the old thread. From what I understood, using > "VMGENID" + "QEMUVGID" worked /well enough/, even if that wasn't > technically in-spec. Ard noted that relying on _CID like that is > technically an ACPI spec notification. So we're between one spec and > another, basically, and doing "VMGENID" + "QEMUVGID" requires fewer > changes, as mentioned, appears to work fine in my testing. > > However, with that said, I think supporting this via "VM_Gen_Counter" > would be a better eventual thing to do, but will require acks and > changes from the ACPI maintainers. Do you think you could prepare your > patch proposal above as something on-top of my tree [1]? And if you can > convince the ACPI maintainers that that's okay, then I'll happily take > the patch. Sure, let me send the ACPI patch stand alone. No need to include the VMGenID change in there. Alex Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
On Fri, Feb 25, 2022 at 03:18:43PM +0100, Alexander Graf wrote: > > I recall this part of the old thread. From what I understood, using > > "VMGENID" + "QEMUVGID" worked /well enough/, even if that wasn't > > technically in-spec. Ard noted that relying on _CID like that is > > technically an ACPI spec notification. So we're between one spec and > > another, basically, and doing "VMGENID" + "QEMUVGID" requires fewer > > changes, as mentioned, appears to work fine in my testing. > > > > However, with that said, I think supporting this via "VM_Gen_Counter" > > would be a better eventual thing to do, but will require acks and > > changes from the ACPI maintainers. Do you think you could prepare your > > patch proposal above as something on-top of my tree [1]? And if you can > > convince the ACPI maintainers that that's okay, then I'll happily take > > the patch. > > > Sure, let me send the ACPI patch stand alone. No need to include the > VMGenID change in there. That's fine. If the ACPI people take it for 5.18, then we can count on it being there and adjust the vmgenid driver accordingly also for 5.18. I just booted up a Windows VM, and it looks like Hyper-V uses "Hyper_V_Gen_Counter_V1", which is also quite long, so we can't really HID match on that either. Jason
On Fri, Feb 25, 2022 at 02:57:38PM +0100, Alexander Graf wrote: > > + > > + phys_addr = (obj->package.elements[0].integer.value << 0) | > > + (obj->package.elements[1].integer.value << 32); > > + state->next_id = devm_memremap(&device->dev, phys_addr, VMGENID_SIZE, MEMREMAP_WB); > > + if (!state->next_id) { > > + ret = -ENOMEM; > > + goto out; > > + } > > + > > + memcpy(state->this_id, state->next_id, sizeof(state->this_id)); > > + add_device_randomness(state->this_id, sizeof(state->this_id)); > > > Please expose the vmgenid via /sysfs so that user space even remotely has a > chance to check if it's been cloned. Export it how? And why, who would care? thanks, greg k-h
Hi Alex, Missed this remark before: On Fri, Feb 25, 2022 at 02:57:38PM +0100, Alexander Graf wrote: > Please expose the vmgenid via /sysfs so that user space even remotely > has a chance to check if it's been cloned. No. Did you read the 0/2 cover letter? I'll quote it for you here: > As a side note, this series intentionally does _not_ focus on > notification of these events to userspace or to other kernel consumers. > Since these VM fork detection events first need to hit the RNG, we can > later talk about what sorts of notifications or mmap'd counters the RNG > should be making accessible to elsewhere. But that's a different sort of > project and ties into a lot of more complicated concerns beyond this > more basic patchset. So hopefully we can keep the discussion rather > focused here to this ACPI business. What about that was unclear to you? Anyway, it's a different thing that will have to be designed and considered carefully, and that design doesn't have a whole lot to do with this little driver here, except insofar as it could build on top of it in one way or another. Yes, it's an important thing to do. No, I'm not going to do it in this patch here. If you want to have a discussion about that, start a different thread. Thanks, Jason
On Fri, 25 Feb 2022 at 14:58, Alexander Graf <graf@amazon.com> wrote: > > > On 25.02.22 13:48, Jason A. Donenfeld wrote: > > > > VM Generation ID is a feature from Microsoft, described at > > <https://go.microsoft.com/fwlink/?LinkId=260709>, and supported by > > Hyper-V and QEMU. Its usage is described in Microsoft's RNG whitepaper, > > <https://aka.ms/win10rng>, as: > > > > If the OS is running in a VM, there is a problem that most > > hypervisors can snapshot the state of the machine and later rewind > > the VM state to the saved state. This results in the machine running > > a second time with the exact same RNG state, which leads to serious > > security problems. To reduce the window of vulnerability, Windows > > 10 on a Hyper-V VM will detect when the VM state is reset, retrieve > > a unique (not random) value from the hypervisor, and reseed the root > > RNG with that unique value. This does not eliminate the > > vulnerability, but it greatly reduces the time during which the RNG > > system will produce the same outputs as it did during a previous > > instantiation of the same VM state. > > > > Linux has the same issue, and given that vmgenid is supported already by > > multiple hypervisors, we can implement more or less the same solution. > > So this commit wires up the vmgenid ACPI notification to the RNG's newly > > added add_vmfork_randomness() function. > > > > It can be used from qemu via the `-device vmgenid,guid=auto` parameter. > > After setting that, use `savevm` in the monitor to save the VM state, > > then quit QEMU, start it again, and use `loadvm`. That will trigger this > > driver's notify function, which hands the new UUID to the RNG. This is > > described in <https://git.qemu.org/?p=qemu.git;a=blob;f=docs/specs/vmgenid.txt>. > > And there are hooks for this in libvirt as well, described in > > <https://libvirt.org/formatdomain.html#general-metadata>. > > > > Note, however, that the treatment of this as a UUID is considered to be > > an accidental QEMU nuance, per > > <https://github.com/libguestfs/virt-v2v/blob/master/docs/vm-generation-id-across-hypervisors.txt>, > > so this driver simply treats these bytes as an opaque 128-bit binary > > blob, as per the spec. This doesn't really make a difference anyway, > > considering that's how it ends up when handed to the RNG in the end. > > > > Cc: Adrian Catangiu <adrian@parity.io> > > Cc: Daniel P. Berrangé <berrange@redhat.com> > > Cc: Dominik Brodowski <linux@dominikbrodowski.net> > > Cc: Ard Biesheuvel <ardb@kernel.org> > > Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> > > Reviewed-by: Laszlo Ersek <lersek@redhat.com> > > Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> > > --- ... > > + > > + device->driver_data = state; > > + > > +out: > > + ACPI_FREE(parsed.pointer); > > + return ret; > > +} > > + > > +static void vmgenid_acpi_notify(struct acpi_device *device, u32 event) > > +{ > > + struct vmgenid_state *state = acpi_driver_data(device); > > + u8 old_id[VMGENID_SIZE]; > > + > > + memcpy(old_id, state->this_id, sizeof(old_id)); > > + memcpy(state->this_id, state->next_id, sizeof(state->this_id)); > > + if (!memcmp(old_id, state->this_id, sizeof(old_id))) > > + return; > > + add_vmfork_randomness(state->this_id, sizeof(state->this_id)); > > +} > > + > > +static const struct acpi_device_id vmgenid_ids[] = { > > + { "VMGENID", 0 }, > > + { "QEMUVGID", 0 }, > > > According to the VMGenID spec[1], you can only rely on _CID and _DDN for > matching. They both contain "VM_Gen_Counter". The list above contains > _HID values which are not an official identifier for the VMGenID device. > > IIRC the ACPI device match logic does match _CID in addition to _HID. > However, it is limited to 8 characters. Let me paste an experimental > hack I did back then to do the _CID matching instead. > > [1] > https://download.microsoft.com/download/3/1/C/31CFC307-98CA-4CA5-914C-D9772691E214/VirtualMachineGenerationID.docx > I think matching on the HIDs of two known existing implementations is fine, as opposed to matching on the (broken) CID of any implementation that claims to be compatible with it. And dumping random strings into the _CID property doesn't mesh well with the ACPI spec either, which is why we don't currently support it. We could still check _DDN if we wanted to, but I don't think this is necessary. Other implementations that want to target his driver explicitly can always put VMGENID or QEMUVGID into the _CID.
On Fri, 25 Feb 2022 at 13:53, Greg KH <gregkh@linuxfoundation.org> wrote: > > On Fri, Feb 25, 2022 at 01:48:48PM +0100, Jason A. Donenfeld wrote: > > +static struct acpi_driver acpi_driver = { > > + .name = "vmgenid", > > + .ids = vmgenid_ids, > > + .owner = THIS_MODULE, > > + .ops = { > > + .add = vmgenid_acpi_add, > > + .notify = vmgenid_acpi_notify, > > + } > > +}; > > + > > +static int __init vmgenid_init(void) > > +{ > > + return acpi_bus_register_driver(&acpi_driver); > > +} > > + > > +static void __exit vmgenid_exit(void) > > +{ > > + acpi_bus_unregister_driver(&acpi_driver); > > +} > > + > > +module_init(vmgenid_init); > > +module_exit(vmgenid_exit); > > Nit, you could use module_acpi_driver() to make this even smaller if you > want to. > With that suggestion adopted, Reviewed-by: Ard Biesheuvel <ardb@kernel.org>
On 25.02.22 15:33, Jason A. Donenfeld wrote: > On Fri, Feb 25, 2022 at 03:18:43PM +0100, Alexander Graf wrote: >>> I recall this part of the old thread. From what I understood, using >>> "VMGENID" + "QEMUVGID" worked /well enough/, even if that wasn't >>> technically in-spec. Ard noted that relying on _CID like that is >>> technically an ACPI spec notification. So we're between one spec and >>> another, basically, and doing "VMGENID" + "QEMUVGID" requires fewer >>> changes, as mentioned, appears to work fine in my testing. >>> >>> However, with that said, I think supporting this via "VM_Gen_Counter" >>> would be a better eventual thing to do, but will require acks and >>> changes from the ACPI maintainers. Do you think you could prepare your >>> patch proposal above as something on-top of my tree [1]? And if you can >>> convince the ACPI maintainers that that's okay, then I'll happily take >>> the patch. >> >> Sure, let me send the ACPI patch stand alone. No need to include the >> VMGenID change in there. > That's fine. If the ACPI people take it for 5.18, then we can count on > it being there and adjust the vmgenid driver accordingly also for 5.18. > > I just booted up a Windows VM, and it looks like Hyper-V uses > "Hyper_V_Gen_Counter_V1", which is also quite long, so we can't really > HID match on that either. Yes, due to the same problem. I'd really prefer we sort out the ACPI matching before this goes mainline. Matching on _HID is explicitly discouraged in the VMGenID spec. Alex Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
On 25.02.22 16:03, Ard Biesheuvel wrote: > On Fri, 25 Feb 2022 at 14:58, Alexander Graf<graf@amazon.com> wrote: >> >> On 25.02.22 13:48, Jason A. Donenfeld wrote: >>> VM Generation ID is a feature from Microsoft, described at >>> <https://go.microsoft.com/fwlink/?LinkId=260709>, and supported by >>> Hyper-V and QEMU. Its usage is described in Microsoft's RNG whitepaper, >>> <https://aka.ms/win10rng>, as: >>> >>> If the OS is running in a VM, there is a problem that most >>> hypervisors can snapshot the state of the machine and later rewind >>> the VM state to the saved state. This results in the machine running >>> a second time with the exact same RNG state, which leads to serious >>> security problems. To reduce the window of vulnerability, Windows >>> 10 on a Hyper-V VM will detect when the VM state is reset, retrieve >>> a unique (not random) value from the hypervisor, and reseed the root >>> RNG with that unique value. This does not eliminate the >>> vulnerability, but it greatly reduces the time during which the RNG >>> system will produce the same outputs as it did during a previous >>> instantiation of the same VM state. >>> >>> Linux has the same issue, and given that vmgenid is supported already by >>> multiple hypervisors, we can implement more or less the same solution. >>> So this commit wires up the vmgenid ACPI notification to the RNG's newly >>> added add_vmfork_randomness() function. >>> >>> It can be used from qemu via the `-device vmgenid,guid=auto` parameter. >>> After setting that, use `savevm` in the monitor to save the VM state, >>> then quit QEMU, start it again, and use `loadvm`. That will trigger this >>> driver's notify function, which hands the new UUID to the RNG. This is >>> described in<https://git.qemu.org/?p=qemu.git;a=blob;f=docs/specs/vmgenid.txt>. >>> And there are hooks for this in libvirt as well, described in >>> <https://libvirt.org/formatdomain.html#general-metadata>. >>> >>> Note, however, that the treatment of this as a UUID is considered to be >>> an accidental QEMU nuance, per >>> <https://github.com/libguestfs/virt-v2v/blob/master/docs/vm-generation-id-across-hypervisors.txt>, >>> so this driver simply treats these bytes as an opaque 128-bit binary >>> blob, as per the spec. This doesn't really make a difference anyway, >>> considering that's how it ends up when handed to the RNG in the end. >>> >>> Cc: Adrian Catangiu<adrian@parity.io> >>> Cc: Daniel P. Berrangé<berrange@redhat.com> >>> Cc: Dominik Brodowski<linux@dominikbrodowski.net> >>> Cc: Ard Biesheuvel<ardb@kernel.org> >>> Cc: Greg Kroah-Hartman<gregkh@linuxfoundation.org> >>> Reviewed-by: Laszlo Ersek<lersek@redhat.com> >>> Signed-off-by: Jason A. Donenfeld<Jason@zx2c4.com> >>> --- > ... >>> + >>> + device->driver_data = state; >>> + >>> +out: >>> + ACPI_FREE(parsed.pointer); >>> + return ret; >>> +} >>> + >>> +static void vmgenid_acpi_notify(struct acpi_device *device, u32 event) >>> +{ >>> + struct vmgenid_state *state = acpi_driver_data(device); >>> + u8 old_id[VMGENID_SIZE]; >>> + >>> + memcpy(old_id, state->this_id, sizeof(old_id)); >>> + memcpy(state->this_id, state->next_id, sizeof(state->this_id)); >>> + if (!memcmp(old_id, state->this_id, sizeof(old_id))) >>> + return; >>> + add_vmfork_randomness(state->this_id, sizeof(state->this_id)); >>> +} >>> + >>> +static const struct acpi_device_id vmgenid_ids[] = { >>> + { "VMGENID", 0 }, >>> + { "QEMUVGID", 0 }, >> >> According to the VMGenID spec[1], you can only rely on _CID and _DDN for >> matching. They both contain "VM_Gen_Counter". The list above contains >> _HID values which are not an official identifier for the VMGenID device. >> >> IIRC the ACPI device match logic does match _CID in addition to _HID. >> However, it is limited to 8 characters. Let me paste an experimental >> hack I did back then to do the _CID matching instead. >> >> [1] >> https://download.microsoft.com/download/3/1/C/31CFC307-98CA-4CA5-914C-D9772691E214/VirtualMachineGenerationID.docx >> > I think matching on the HIDs of two known existing implementations is > fine, as opposed to matching on the (broken) CID of any implementation > that claims to be compatible with it. And dumping random strings into > the _CID property doesn't mesh well with the ACPI spec either, which > is why we don't currently support it. > > We could still check _DDN if we wanted to, but I don't think this is > necessary. Other implementations that want to target his driver > explicitly can always put VMGENID or QEMUVGID into the _CID. No, they can not. The spec explicitly also states that _HID has to be unique per hypervisor: "A hypervisor must also expose an object below the device, named "_HID", which is the "hardware ID." This value is hypervisor specific and must be different for each vendor. The Windows generation ID guest driver will load on the compatible ID of "VM_Gen_Counter", so it will be present regardless of which hypervisor it is running on top of, as long as that hypervisor follows the rules above." Alex Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
On 25.02.22 15:54, Jason A. Donenfeld wrote: > Hi Alex, > > Missed this remark before: > > On Fri, Feb 25, 2022 at 02:57:38PM +0100, Alexander Graf wrote: >> Please expose the vmgenid via /sysfs so that user space even remotely >> has a chance to check if it's been cloned. > No. Did you read the 0/2 cover letter? I'll quote it for you here: > >> As a side note, this series intentionally does _not_ focus on >> notification of these events to userspace or to other kernel consumers. >> Since these VM fork detection events first need to hit the RNG, we can >> later talk about what sorts of notifications or mmap'd counters the RNG >> should be making accessible to elsewhere. But that's a different sort of >> project and ties into a lot of more complicated concerns beyond this >> more basic patchset. So hopefully we can keep the discussion rather >> focused here to this ACPI business. > What about that was unclear to you? > > Anyway, it's a different thing that will have to be designed and > considered carefully, and that design doesn't have a whole lot to do > with this little driver here, except insofar as it could build on top of > it in one way or another. Yes, it's an important thing to do. No, I'm > not going to do it in this patch here. If you want to have a discussion > about that, start a different thread. I'm not talking about a notification interface - we've gone through great length on that one in the previous submission. What I'm more interested in is *any* way for user space to read the current VM Gen ID. The same way I'm interested to see other device attributes of my system through sysfs. Alex Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
On Fri, 25 Feb 2022 at 16:12, Alexander Graf <graf@amazon.com> wrote: > > > On 25.02.22 15:33, Jason A. Donenfeld wrote: > > On Fri, Feb 25, 2022 at 03:18:43PM +0100, Alexander Graf wrote: > >>> I recall this part of the old thread. From what I understood, using > >>> "VMGENID" + "QEMUVGID" worked /well enough/, even if that wasn't > >>> technically in-spec. Ard noted that relying on _CID like that is > >>> technically an ACPI spec notification. So we're between one spec and > >>> another, basically, and doing "VMGENID" + "QEMUVGID" requires fewer > >>> changes, as mentioned, appears to work fine in my testing. > >>> > >>> However, with that said, I think supporting this via "VM_Gen_Counter" > >>> would be a better eventual thing to do, but will require acks and > >>> changes from the ACPI maintainers. Do you think you could prepare your > >>> patch proposal above as something on-top of my tree [1]? And if you can > >>> convince the ACPI maintainers that that's okay, then I'll happily take > >>> the patch. > >> > >> Sure, let me send the ACPI patch stand alone. No need to include the > >> VMGenID change in there. > > That's fine. If the ACPI people take it for 5.18, then we can count on > > it being there and adjust the vmgenid driver accordingly also for 5.18. > > > > I just booted up a Windows VM, and it looks like Hyper-V uses > > "Hyper_V_Gen_Counter_V1", which is also quite long, so we can't really > > HID match on that either. > > > Yes, due to the same problem. I'd really prefer we sort out the ACPI > matching before this goes mainline. Matching on _HID is explicitly > discouraged in the VMGenID spec. > OK, this really sucks. Quoting the ACPI spec: """ A _HID object evaluates to either a numeric 32-bit compressed EISA type ID or a string. If a string, the format must be an alphanumeric PNP or ACPI ID with no asterisk or other leading characters. A valid PNP ID must be of the form "AAA####" where A is an uppercase letter and # is a hex digit. A valid ACPI ID must be of the form "NNNN####" where N is an uppercase letter or a digit ('0'-'9') and # is a hex digit. This specification reserves the string "ACPI" for use only with devices defined herein. It further reserves all strings representing 4 HEX digits for exclusive use with PCI-assigned Vendor IDs. """ So now we have to implement Microsoft's fork of ACPI to be able to use this device, even if we expose it from QEMU instead of Hyper-V? I strongly object to that. Instead, we can match on _HID exposed by QEMU, and cordially invite Microsoft to align their spec with the ACPI spec.
On 25.02.22 16:16, Ard Biesheuvel wrote: > On Fri, 25 Feb 2022 at 16:12, Alexander Graf <graf@amazon.com> wrote: >> >> On 25.02.22 15:33, Jason A. Donenfeld wrote: >>> On Fri, Feb 25, 2022 at 03:18:43PM +0100, Alexander Graf wrote: >>>>> I recall this part of the old thread. From what I understood, using >>>>> "VMGENID" + "QEMUVGID" worked /well enough/, even if that wasn't >>>>> technically in-spec. Ard noted that relying on _CID like that is >>>>> technically an ACPI spec notification. So we're between one spec and >>>>> another, basically, and doing "VMGENID" + "QEMUVGID" requires fewer >>>>> changes, as mentioned, appears to work fine in my testing. >>>>> >>>>> However, with that said, I think supporting this via "VM_Gen_Counter" >>>>> would be a better eventual thing to do, but will require acks and >>>>> changes from the ACPI maintainers. Do you think you could prepare your >>>>> patch proposal above as something on-top of my tree [1]? And if you can >>>>> convince the ACPI maintainers that that's okay, then I'll happily take >>>>> the patch. >>>> Sure, let me send the ACPI patch stand alone. No need to include the >>>> VMGenID change in there. >>> That's fine. If the ACPI people take it for 5.18, then we can count on >>> it being there and adjust the vmgenid driver accordingly also for 5.18. >>> >>> I just booted up a Windows VM, and it looks like Hyper-V uses >>> "Hyper_V_Gen_Counter_V1", which is also quite long, so we can't really >>> HID match on that either. >> >> Yes, due to the same problem. I'd really prefer we sort out the ACPI >> matching before this goes mainline. Matching on _HID is explicitly >> discouraged in the VMGenID spec. >> > OK, this really sucks. Quoting the ACPI spec: > > """ > A _HID object evaluates to either a numeric 32-bit compressed EISA > type ID or a string. If a string, the format must be an alphanumeric > PNP or ACPI ID with no asterisk or other leading characters. > A valid PNP ID must be of the form "AAA####" where A is an uppercase > letter and # is a hex digit. > A valid ACPI ID must be of the form "NNNN####" where N is an uppercase > letter or a digit ('0'-'9') and # is a hex digit. This specification > reserves the string "ACPI" for use only with devices defined herein. > It further reserves all strings representing 4 HEX digits for > exclusive use with PCI-assigned Vendor IDs. > """ > > So now we have to implement Microsoft's fork of ACPI to be able to use > this device, even if we expose it from QEMU instead of Hyper-V? I > strongly object to that. > > Instead, we can match on _HID exposed by QEMU, and cordially invite > Microsoft to align their spec with the ACPI spec. Doing that would be a backwards incompatible change for Hyper-V, no? I understand that you're upset about their spec, but that doesn't mean we can't find a path forward to make it all compatible. IMHO just matching on the first 9 bytes of the _CID/_HID string is perfectly fine. It follows the spec, but still allows for weird identifiers like this one to work. I don't understand the rush here. This had been sitting on the ML for 1 year - and now suddenly talking the match through properly and getting VMGenID spec compatible matching support into the ACPI core is a problem? What did I miss? :) Alex Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
Hi Alex, On Fri, Feb 25, 2022 at 04:15:59PM +0100, Alexander Graf wrote: > I'm not talking about a notification interface - we've gone through > great length on that one in the previous submission. What I'm more > interested in is *any* way for user space to read the current VM Gen ID. > The same way I'm interested to see other device attributes of my system > through sysfs. Again, no. Same basic objection: we can do this later and design it coherently with the rest. For example, maybe it's better to expose a generation counter rather than 16 byte blob, and expect userspace to call getrandom() subsequently to get something fresh. Or not! But maybe it should be hashed with a fixed prefix string before being exposed to userspace. Or not! I don't know, but that's not going to happen on this patchset. There is no reason at all why that needs to be done here and now. Trying to do too much at the same time is likely why the previous efforts from your team stalled out last year. Propose something later, in a new thread, and we can discuss then. One step at a time... Jason
On 25.02.22 15:36, Greg KH wrote: > On Fri, Feb 25, 2022 at 02:57:38PM +0100, Alexander Graf wrote: >>> + >>> + phys_addr = (obj->package.elements[0].integer.value << 0) | >>> + (obj->package.elements[1].integer.value << 32); >>> + state->next_id = devm_memremap(&device->dev, phys_addr, VMGENID_SIZE, MEMREMAP_WB); >>> + if (!state->next_id) { >>> + ret = -ENOMEM; >>> + goto out; >>> + } >>> + >>> + memcpy(state->this_id, state->next_id, sizeof(state->this_id)); >>> + add_device_randomness(state->this_id, sizeof(state->this_id)); >> >> Please expose the vmgenid via /sysfs so that user space even remotely has a >> chance to check if it's been cloned. > Export it how? And why, who would care? You can just create a sysfs file that contains it. The same way we have sysfs files for UEFI config tables. Or sysfs files for the acpi device nodes themselves. I personally don't care if we put this into a generic location (/sys/hypervisor for example) or into the existing acpi device node as additional file you can just read. Who would care? Well, for starters I would care for debugging purposes :). Extracting the ID and validating that it's different than before is quite useful when you want to check if the clone rng adjustment actually worked. I don't have very strong feelings on it though - unlike the _CID conversation. Alex Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
On Fri, Feb 25, 2022 at 04:16:27PM +0100, Ard Biesheuvel wrote: > > > I just booted up a Windows VM, and it looks like Hyper-V uses > > > "Hyper_V_Gen_Counter_V1", which is also quite long, so we can't really > > > HID match on that either. > > > > > > Yes, due to the same problem. I'd really prefer we sort out the ACPI > > matching before this goes mainline. Matching on _HID is explicitly > > discouraged in the VMGenID spec. > > > > OK, this really sucks. Quoting the ACPI spec: > > """ > A _HID object evaluates to either a numeric 32-bit compressed EISA > type ID or a string. If a string, the format must be an alphanumeric > PNP or ACPI ID with no asterisk or other leading characters. > A valid PNP ID must be of the form "AAA####" where A is an uppercase > letter and # is a hex digit. > A valid ACPI ID must be of the form "NNNN####" where N is an uppercase > letter or a digit ('0'-'9') and # is a hex digit. This specification > reserves the string "ACPI" for use only with devices defined herein. > It further reserves all strings representing 4 HEX digits for > exclusive use with PCI-assigned Vendor IDs. > """ > > So now we have to implement Microsoft's fork of ACPI to be able to use > this device, even if we expose it from QEMU instead of Hyper-V? I > strongly object to that. > > Instead, we can match on _HID exposed by QEMU, and cordially invite > Microsoft to align their spec with the ACPI spec. I don't know about that... Seems a bit extreme. Hopefully Alex will be able to sort something out with the ACPI people, and this driver will work inside of Hyper-V. Here's what we currently have: static const struct acpi_device_id vmgenid_ids[] = { { "VMGENID", 0 }, <------------------------------------ ??? { "QEMUVGID", 0 }, <------------------------------------ QEMU { }, }; Adrian added "VMGENID" in last year's v4, so I copied that for this new driver here. But does anybody know which hypervisor it is for? Some internal Amazon thing? Firecracker? VMware? In case Alex does not succeed with the ACPI changes, it'd be nice to know which HIDs for which hypervisors we do and do not support. Jason
Hi again, On Fri, Feb 25, 2022 at 04:31:02PM +0100, Alexander Graf wrote: > >> Please expose the vmgenid via /sysfs so that user space even remotely has a > >> chance to check if it's been cloned. > > Export it how? And why, who would care? > You can just As mentioned in <https://lore.kernel.org/lkml/Yhj1nYHXmimPsqFd@zx2c4.com/>, propose something later for all of this. It doesn't need to happen all at once *now*. Thanks, Jason
On 25.02.22 16:34, Jason A. Donenfeld wrote: > On Fri, Feb 25, 2022 at 04:16:27PM +0100, Ard Biesheuvel wrote: >>>> I just booted up a Windows VM, and it looks like Hyper-V uses >>>> "Hyper_V_Gen_Counter_V1", which is also quite long, so we can't really >>>> HID match on that either. >>> >>> Yes, due to the same problem. I'd really prefer we sort out the ACPI >>> matching before this goes mainline. Matching on _HID is explicitly >>> discouraged in the VMGenID spec. >>> >> OK, this really sucks. Quoting the ACPI spec: >> >> """ >> A _HID object evaluates to either a numeric 32-bit compressed EISA >> type ID or a string. If a string, the format must be an alphanumeric >> PNP or ACPI ID with no asterisk or other leading characters. >> A valid PNP ID must be of the form "AAA####" where A is an uppercase >> letter and # is a hex digit. >> A valid ACPI ID must be of the form "NNNN####" where N is an uppercase >> letter or a digit ('0'-'9') and # is a hex digit. This specification >> reserves the string "ACPI" for use only with devices defined herein. >> It further reserves all strings representing 4 HEX digits for >> exclusive use with PCI-assigned Vendor IDs. >> """ >> >> So now we have to implement Microsoft's fork of ACPI to be able to use >> this device, even if we expose it from QEMU instead of Hyper-V? I >> strongly object to that. >> >> Instead, we can match on _HID exposed by QEMU, and cordially invite >> Microsoft to align their spec with the ACPI spec. > I don't know about that... Seems a bit extreme. Hopefully Alex will be > able to sort something out with the ACPI people, and this driver will > work inside of Hyper-V. > > Here's what we currently have: > > static const struct acpi_device_id vmgenid_ids[] = { > { "VMGENID", 0 }, <------------------------------------ ??? > { "QEMUVGID", 0 }, <------------------------------------ QEMU > { }, > }; > > Adrian added "VMGENID" in last year's v4, so I copied that for this new > driver here. But does anybody know which hypervisor it is for? Some > internal Amazon thing? Firecracker? VMware? In case Alex does not > succeed with the ACPI changes, it'd be nice to know which HIDs for > which hypervisors we do and do not support. I believe "VMGENID" was for the firecracker prototype that Adrian built back then, yeah. Matching on _HID for this is a rat hole unfortunately, so let's see what the ACPI patch gets us :). Alex Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
Hi Alex, On Fri, Feb 25, 2022 at 04:22:54PM +0100, Alexander Graf wrote: > I don't understand the rush here. This had been sitting on the ML for 1 > year - and now suddenly talking the match through properly and getting > VMGenID spec compatible matching support into the ACPI core is a > problem? What did I miss? :) I don't think this is a question about speed. Ard doesn't like the spec. You like the feature more than you dislike the spec. Apparently that means there's a disagreement. As I mentioned earlier, I'd encourage you to send a patch to the ACPI people and let them decide. If that gets in, then I'm fine with modifying vmgenid to meet the spec and take advantage of the change you'll be making to the ACPI code. If it is rejected by the ACPI people, and consequently Linux isn't able to match on _CID, then I guess we'll have the next best thing, which is "well, it still works on QEMU." Hopefully you'll convince them. Feel free to CC me on that patch. Jason
Hi Alex, On Fri, Feb 25, 2022 at 04:37:43PM +0100, Alexander Graf wrote: > I believe "VMGENID" was for the firecracker prototype that Adrian built > back then, yeah. Matching on _HID for this is a rat hole unfortunately, > so let's see what the ACPI patch gets us :). Thanks. I'll add a comment to the code about Firecracker. And indeed hopefully that'll all go away anyway. Jason
On 25.02.22 16:43, Jason A. Donenfeld wrote: > > Hi Alex, > > On Fri, Feb 25, 2022 at 04:22:54PM +0100, Alexander Graf wrote: >> I don't understand the rush here. This had been sitting on the ML for 1 >> year - and now suddenly talking the match through properly and getting >> VMGenID spec compatible matching support into the ACPI core is a >> problem? What did I miss? :) > I don't think this is a question about speed. Ard doesn't like the spec. > You like the feature more than you dislike the spec. Apparently that > means there's a disagreement. > > As I mentioned earlier, I'd encourage you to send a patch to the ACPI > people and let them decide. If that gets in, then I'm fine with > modifying vmgenid to meet the spec and take advantage of the change > you'll be making to the ACPI code. If it is rejected by the ACPI people, > and consequently Linux isn't able to match on _CID, then I guess we'll > have the next best thing, which is "well, it still works on QEMU." > Hopefully you'll convince them. Feel free to CC me on that patch. I agree with that approach and CC'ed you on the ACPI patch :). Let's explore all options to match against _CID before we give up. Alex Amazon Development Center Germany GmbH Krausenstr. 38 10117 Berlin Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B Sitz: Berlin Ust-ID: DE 289 237 879
diff --git a/MAINTAINERS b/MAINTAINERS index 777cd6fa2b3d..a10997e15146 100644 --- a/MAINTAINERS +++ b/MAINTAINERS @@ -16211,6 +16211,7 @@ M: Jason A. Donenfeld <Jason@zx2c4.com> T: git https://git.kernel.org/pub/scm/linux/kernel/git/crng/random.git S: Maintained F: drivers/char/random.c +F: drivers/virt/vmgenid.c RAPIDIO SUBSYSTEM M: Matt Porter <mporter@kernel.crashing.org> diff --git a/drivers/virt/Kconfig b/drivers/virt/Kconfig index 8061e8ef449f..5596c7313f59 100644 --- a/drivers/virt/Kconfig +++ b/drivers/virt/Kconfig @@ -13,6 +13,15 @@ menuconfig VIRT_DRIVERS if VIRT_DRIVERS +config VMGENID + tristate "Virtual Machine Generation ID driver" + default m + depends on ACPI + help + Say Y here to use the hypervisor-provided Virtual Machine Generation ID + to reseed the RNG when the VM is cloned. This is highly recommended if + you intend to do any rollback / cloning / snapshotting of VMs. + config FSL_HV_MANAGER tristate "Freescale hypervisor management driver" depends on FSL_SOC diff --git a/drivers/virt/Makefile b/drivers/virt/Makefile index 3e272ea60cd9..108d0ffcc9aa 100644 --- a/drivers/virt/Makefile +++ b/drivers/virt/Makefile @@ -4,6 +4,7 @@ # obj-$(CONFIG_FSL_HV_MANAGER) += fsl_hypervisor.o +obj-$(CONFIG_VMGENID) += vmgenid.o obj-y += vboxguest/ obj-$(CONFIG_NITRO_ENCLAVES) += nitro_enclaves/ diff --git a/drivers/virt/vmgenid.c b/drivers/virt/vmgenid.c new file mode 100644 index 000000000000..e3dd4afb33c6 --- /dev/null +++ b/drivers/virt/vmgenid.c @@ -0,0 +1,112 @@ +// SPDX-License-Identifier: GPL-2.0 +/* + * Copyright (C) 2022 Jason A. Donenfeld <Jason@zx2c4.com>. All Rights Reserved. + * + * The "Virtual Machine Generation ID" is exposed via ACPI and changes when a + * virtual machine forks or is cloned. This driver exists for shepherding that + * information to random.c. + */ + +#include <linux/kernel.h> +#include <linux/module.h> +#include <linux/acpi.h> +#include <linux/random.h> + +ACPI_MODULE_NAME("vmgenid"); + +enum { VMGENID_SIZE = 16 }; + +struct vmgenid_state { + u8 *next_id; + u8 this_id[VMGENID_SIZE]; +}; + +static int vmgenid_acpi_add(struct acpi_device *device) +{ + struct acpi_buffer parsed = { ACPI_ALLOCATE_BUFFER }; + struct vmgenid_state *state; + union acpi_object *obj; + phys_addr_t phys_addr; + acpi_status status; + int ret = 0; + + state = devm_kmalloc(&device->dev, sizeof(*state), GFP_KERNEL); + if (!state) + return -ENOMEM; + + status = acpi_evaluate_object(device->handle, "ADDR", NULL, &parsed); + if (ACPI_FAILURE(status)) { + ACPI_EXCEPTION((AE_INFO, status, "Evaluating ADDR")); + return -ENODEV; + } + obj = parsed.pointer; + if (!obj || obj->type != ACPI_TYPE_PACKAGE || obj->package.count != 2 || + obj->package.elements[0].type != ACPI_TYPE_INTEGER || + obj->package.elements[1].type != ACPI_TYPE_INTEGER) { + ret = -EINVAL; + goto out; + } + + phys_addr = (obj->package.elements[0].integer.value << 0) | + (obj->package.elements[1].integer.value << 32); + state->next_id = devm_memremap(&device->dev, phys_addr, VMGENID_SIZE, MEMREMAP_WB); + if (!state->next_id) { + ret = -ENOMEM; + goto out; + } + + memcpy(state->this_id, state->next_id, sizeof(state->this_id)); + add_device_randomness(state->this_id, sizeof(state->this_id)); + + device->driver_data = state; + +out: + ACPI_FREE(parsed.pointer); + return ret; +} + +static void vmgenid_acpi_notify(struct acpi_device *device, u32 event) +{ + struct vmgenid_state *state = acpi_driver_data(device); + u8 old_id[VMGENID_SIZE]; + + memcpy(old_id, state->this_id, sizeof(old_id)); + memcpy(state->this_id, state->next_id, sizeof(state->this_id)); + if (!memcmp(old_id, state->this_id, sizeof(old_id))) + return; + add_vmfork_randomness(state->this_id, sizeof(state->this_id)); +} + +static const struct acpi_device_id vmgenid_ids[] = { + { "VMGENID", 0 }, + { "QEMUVGID", 0 }, + { }, +}; + +static struct acpi_driver acpi_driver = { + .name = "vmgenid", + .ids = vmgenid_ids, + .owner = THIS_MODULE, + .ops = { + .add = vmgenid_acpi_add, + .notify = vmgenid_acpi_notify, + } +}; + +static int __init vmgenid_init(void) +{ + return acpi_bus_register_driver(&acpi_driver); +} + +static void __exit vmgenid_exit(void) +{ + acpi_bus_unregister_driver(&acpi_driver); +} + +module_init(vmgenid_init); +module_exit(vmgenid_exit); + +MODULE_DEVICE_TABLE(acpi, vmgenid_ids); +MODULE_DESCRIPTION("Virtual Machine Generation ID"); +MODULE_LICENSE("GPL v2"); +MODULE_AUTHOR("Jason A. Donenfeld <Jason@zx2c4.com>");
VM Generation ID is a feature from Microsoft, described at <https://go.microsoft.com/fwlink/?LinkId=260709>, and supported by Hyper-V and QEMU. Its usage is described in Microsoft's RNG whitepaper, <https://aka.ms/win10rng>, as: If the OS is running in a VM, there is a problem that most hypervisors can snapshot the state of the machine and later rewind the VM state to the saved state. This results in the machine running a second time with the exact same RNG state, which leads to serious security problems. To reduce the window of vulnerability, Windows 10 on a Hyper-V VM will detect when the VM state is reset, retrieve a unique (not random) value from the hypervisor, and reseed the root RNG with that unique value. This does not eliminate the vulnerability, but it greatly reduces the time during which the RNG system will produce the same outputs as it did during a previous instantiation of the same VM state. Linux has the same issue, and given that vmgenid is supported already by multiple hypervisors, we can implement more or less the same solution. So this commit wires up the vmgenid ACPI notification to the RNG's newly added add_vmfork_randomness() function. It can be used from qemu via the `-device vmgenid,guid=auto` parameter. After setting that, use `savevm` in the monitor to save the VM state, then quit QEMU, start it again, and use `loadvm`. That will trigger this driver's notify function, which hands the new UUID to the RNG. This is described in <https://git.qemu.org/?p=qemu.git;a=blob;f=docs/specs/vmgenid.txt>. And there are hooks for this in libvirt as well, described in <https://libvirt.org/formatdomain.html#general-metadata>. Note, however, that the treatment of this as a UUID is considered to be an accidental QEMU nuance, per <https://github.com/libguestfs/virt-v2v/blob/master/docs/vm-generation-id-across-hypervisors.txt>, so this driver simply treats these bytes as an opaque 128-bit binary blob, as per the spec. This doesn't really make a difference anyway, considering that's how it ends up when handed to the RNG in the end. Cc: Adrian Catangiu <adrian@parity.io> Cc: Daniel P. Berrangé <berrange@redhat.com> Cc: Dominik Brodowski <linux@dominikbrodowski.net> Cc: Ard Biesheuvel <ardb@kernel.org> Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org> Reviewed-by: Laszlo Ersek <lersek@redhat.com> Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com> --- Changes v3->v4: - Add this driver to MAINTAINERS, per Ard's request. Note: I didn't really want to do this at first, because I was hoping the original Amazon team looking into this last year would step up. But it seems like that team has moved on, and anyway I've basically rewritten the driver from scratch at this point -- not a single line of the original exists -- and so I guess I'll maintain it myself. Adding Greg to the CC for his ack on this. - Don't use a static global state in case there are multiple instances. - Use devm_memremap instead of the acpi internal functions. - Default to being modular instead of a built-in, as apparently this is udev-able. MAINTAINERS | 1 + drivers/virt/Kconfig | 9 ++++ drivers/virt/Makefile | 1 + drivers/virt/vmgenid.c | 112 +++++++++++++++++++++++++++++++++++++++++ 4 files changed, 123 insertions(+) create mode 100644 drivers/virt/vmgenid.c