Changes since v1:
- Switch from Keyring to sysfs (James)
An attestation report is signed evidence of how a Trusted Virtual
Machine (TVM) was launched and its current state. A verifying party uses
the report to make judgements of the confidentiality and integrity of
that execution environment. Upon successful attestation the verifying
party may, for example, proceed to deploy secrets to the TVM to carry
out a workload. Multiple confidential computing platforms share this
similar flow.
The approach of adding adding new char devs and new ioctls, for what
amounts to the same logical functionality with minor formatting
differences across vendors [1], is untenable. Common concepts and the
community benefit from common infrastructure.
Use sysfs for this facility for maintainability compared to ioctl(). The
expectation is that this interface is a boot time, configure once, get
report, and done flow. I.e. not something that receives ongoing
transactions at runtime. However, runtime retrieval is not precluded and
a mechanism to detect potential configuration conflicts from
multiple-threads using this interface is included.
The keyring@ list is dropped on this posting since a new key-type is no
longer being pursued.
Link: http://lore.kernel.org/r/[email protected]
---
Dan Williams (5):
virt: coco: Add a coco/Makefile and coco/Kconfig
tsm: Introduce a shared ABI for attestation reports
virt: sevguest: Prep for kernel internal {get,get_ext}_report()
mm/slab: Add __free() support for kvfree
virt: sevguest: Add TSM_REPORTS support for SNP_{GET,GET_EXT}_REPORT
Documentation/ABI/testing/sysfs-class-tsm | 47 +++++
MAINTAINERS | 8 +
drivers/virt/Kconfig | 6 -
drivers/virt/Makefile | 4
drivers/virt/coco/Kconfig | 13 +
drivers/virt/coco/Makefile | 8 +
drivers/virt/coco/sev-guest/Kconfig | 1
drivers/virt/coco/sev-guest/sev-guest.c | 129 ++++++++++++-
drivers/virt/coco/tdx-guest/Kconfig | 1
drivers/virt/coco/tsm.c | 290 +++++++++++++++++++++++++++++
include/linux/slab.h | 2
include/linux/tsm.h | 45 +++++
12 files changed, 535 insertions(+), 19 deletions(-)
create mode 100644 Documentation/ABI/testing/sysfs-class-tsm
create mode 100644 drivers/virt/coco/Kconfig
create mode 100644 drivers/virt/coco/Makefile
create mode 100644 drivers/virt/coco/tsm.c
create mode 100644 include/linux/tsm.h
base-commit: 06c2afb862f9da8dc5efa4b6076a0e48c3fbaaa5
One of the common operations of a TSM (Trusted Security Module) is to
provide a way for a TVM (confidential computing guest execution
environment) to take a measurement of its launch state, sign it and
submit it to a verifying party. Upon successful attestation that
verifies the integrity of the TVM additional secrets may be deployed.
The concept is common across TSMs, but the implementations are
unfortunately vendor specific. While the industry grapples with a common
definition of this attestation format [1], Linux need not make this
problem worse by defining a new ABI per TSM that wants to perform a
similar operation. The current momentum has been to invent new ioctl-ABI
per TSM per function which at best is an abdication of the kernel's
responsibility to make common infrastructure concepts share common ABI.
The proposal, targeted to conceptually work with TDX, SEV, COVE if not
more, is to define a sysfs interface to retrieve the TSM-specific blob.
echo $hex_encoded_userdata_plus_nonce > /sys/class/tsm/tsm0/inhex
hexdump /sys/class/tsm/tsm0/outblob
This approach later allows for the standardization of the attestation
blob format without needing to change the Linux ABI. Until then, the
format of 'outblob' is determined by the parent device for 'tsm0'.
The expectation is that this is a boot time exchange that need not be
regenerated, making it amenable to a sysfs interface. In case userspace
does try to generate multiple attestation reports it includes conflict
detection so userspace can be sure no other thread changed the
parameters from its last configuration step to the blob retrieval.
TSM specific options are encoded as 'extra' attributes on the TSM device
with the expectation that vendors reuse the same options for similar
concepts. The current options are defined by SEV-SNP's need for a
'privilege level' concept (VMPL), and the option to retrieve a
certificate chain in addition to the attestation report ("extended"
format).
Link: http://lore.kernel.org/r/[email protected] [1]
Cc: Kuppuswamy Sathyanarayanan <[email protected]>
Cc: Dionna Amalie Glaze <[email protected]>
Cc: James Bottomley <[email protected]>
Cc: Peter Gonda <[email protected]>
Cc: Greg Kroah-Hartman <[email protected]>
Cc: Samuel Ortiz <[email protected]>
Signed-off-by: Dan Williams <[email protected]>
---
Documentation/ABI/testing/sysfs-class-tsm | 47 +++++
MAINTAINERS | 8 +
drivers/virt/coco/Kconfig | 4
drivers/virt/coco/Makefile | 1
drivers/virt/coco/tdx-guest/Kconfig | 1
drivers/virt/coco/tsm.c | 290 +++++++++++++++++++++++++++++
include/linux/tsm.h | 45 +++++
7 files changed, 396 insertions(+)
create mode 100644 Documentation/ABI/testing/sysfs-class-tsm
create mode 100644 drivers/virt/coco/tsm.c
create mode 100644 include/linux/tsm.h
diff --git a/Documentation/ABI/testing/sysfs-class-tsm b/Documentation/ABI/testing/sysfs-class-tsm
new file mode 100644
index 000000000000..37017bde626d
--- /dev/null
+++ b/Documentation/ABI/testing/sysfs-class-tsm
@@ -0,0 +1,47 @@
+What: /sys/class/tsm/tsm0/inhex
+Date: August, 2023
+KernelVersion: v6.6
+Contact: [email protected]
+Description:
+ (RW) Hex encoded userdata to be included in the attestation
+ report. For replay protection this should include a nonce, but
+ the kernel does not place any restrictions on the content.
+
+What: /sys/class/tsm/tsm0/outblob
+Date: August, 2023
+KernelVersion: v6.6
+Contact: [email protected]
+Description:
+ (RO) Binary attestation report generated from @inhex translated
+ to binary and any options. The format of the report is vendor
+ specific and determined by the parent device of 'tsm0'.
+
+What: /sys/class/tsm/tsm0/generation
+Date: August, 2023
+KernelVersion: v6.6
+Contact: [email protected]
+Description:
+ (RO) The value in this attribute increments each time @inhex or
+ any option is written. Userspace can detect conflicts by
+ checking generation before writing to any attribute and making
+ sure the number of writes matches expectations after reading
+ @outblob.
+
+What: /sys/class/tsm/tsm0/privlevel
+Date: August, 2023
+KernelVersion: v6.6
+Contact: [email protected]
+Description:
+ (RW) If a TSM implementation supports the concept of attestation
+ reports for TVMs running at different privilege levels, like
+ SEV-SNP "VMPL", specify the privilege level via this attribute.
+
+What: /sys/class/tsm/tsm0/format
+Date: August, 2023
+KernelVersion: v6.6
+Contact: [email protected]
+Description:
+ (RW) If a TSM implementation supports the concept of attestation
+ reports with "extended" contents, like SEV-SNP extended reports
+ with certificate chains, specify "extended" vs "default" via
+ this attribute.
diff --git a/MAINTAINERS b/MAINTAINERS
index 3be1bdfe8ecc..97f74d344c8a 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -21625,6 +21625,14 @@ W: https://github.com/srcres258/linux-doc
T: git git://github.com/srcres258/linux-doc.git doc-zh-tw
F: Documentation/translations/zh_TW/
+TRUSTED SECURITY MODULE (TSM) ATTESTATION REPORTS
+M: Dan Williams <[email protected]>
+L: [email protected]
+S: Maintained
+F: Documentation/ABI/testing/sysfs-class-tsm
+F: drivers/virt/coco/tsm.c
+F: include/linux/tsm.h
+
TTY LAYER
M: Greg Kroah-Hartman <[email protected]>
M: Jiri Slaby <[email protected]>
diff --git a/drivers/virt/coco/Kconfig b/drivers/virt/coco/Kconfig
index fc5c64f04c4a..d92f07019f38 100644
--- a/drivers/virt/coco/Kconfig
+++ b/drivers/virt/coco/Kconfig
@@ -2,6 +2,10 @@
#
# Confidential computing related collateral
#
+
+config TSM_REPORTS
+ tristate
+
source "drivers/virt/coco/efi_secret/Kconfig"
source "drivers/virt/coco/sev-guest/Kconfig"
diff --git a/drivers/virt/coco/Makefile b/drivers/virt/coco/Makefile
index 55302ef719ad..18c1aba5edb7 100644
--- a/drivers/virt/coco/Makefile
+++ b/drivers/virt/coco/Makefile
@@ -2,6 +2,7 @@
#
# Confidential computing related collateral
#
+obj-$(CONFIG_TSM_REPORTS) += tsm.o
obj-$(CONFIG_EFI_SECRET) += efi_secret/
obj-$(CONFIG_SEV_GUEST) += sev-guest/
obj-$(CONFIG_INTEL_TDX_GUEST) += tdx-guest/
diff --git a/drivers/virt/coco/tdx-guest/Kconfig b/drivers/virt/coco/tdx-guest/Kconfig
index 14246fc2fb02..22dd59e19431 100644
--- a/drivers/virt/coco/tdx-guest/Kconfig
+++ b/drivers/virt/coco/tdx-guest/Kconfig
@@ -1,6 +1,7 @@
config TDX_GUEST_DRIVER
tristate "TDX Guest driver"
depends on INTEL_TDX_GUEST
+ select TSM_REPORTS
help
The driver provides userspace interface to communicate with
the TDX module to request the TDX guest details like attestation
diff --git a/drivers/virt/coco/tsm.c b/drivers/virt/coco/tsm.c
new file mode 100644
index 000000000000..1bf2ee82eb94
--- /dev/null
+++ b/drivers/virt/coco/tsm.c
@@ -0,0 +1,290 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/* Copyright(c) 2023 Intel Corporation. All rights reserved. */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/tsm.h>
+#include <linux/err.h>
+#include <linux/slab.h>
+#include <linux/rwsem.h>
+#include <linux/device.h>
+#include <linux/string.h>
+#include <linux/module.h>
+#include <linux/cleanup.h>
+
+struct class *tsm_class;
+static struct tsm_provider {
+ const struct tsm_ops *ops;
+ struct device *dev;
+} provider;
+static DECLARE_RWSEM(tsm_rwsem);
+
+/**
+ * DOC: Trusted Security Module (TSM) Attestation Report Interface
+ *
+ * The TSM report interface is a common provider of blobs that facilitate
+ * attestation of a TVM (confidential computing guest) by an attestation
+ * service. A TSM report combines a user-defined blob (likely a public-key with
+ * a nonce for a key-exchange protocol) with a signed attestation report. That
+ * combined blob is then used to obtain secrets provided by an agent that can
+ * validate the attestation report. The expectation is that this interface is
+ * invoked infrequently, likely only once at TVM boot time.
+ *
+ * The attestation report format is TSM provider specific, when / if a standard
+ * materializes that can be published instead of the vendor layout.
+ */
+
+/**
+ * struct tsm_report - track state of report generation relative to options
+ * @desc: report generation options / cached report state
+ * @outblob: generated evidence to provider to the attestation agent
+ * @outblob_len: sizeof(outblob)
+ * @write_generation: conflict detection, and report regeneration tracking
+ * @read_generation: cached report invalidation tracking
+ */
+struct tsm_report {
+ struct tsm_desc desc;
+ size_t outblob_len;
+ u8 *outblob;
+ unsigned long write_generation;
+ unsigned long read_generation;
+} tsm_report;
+
+static ssize_t privlevel_store(struct device *dev,
+ struct device_attribute *attr, const char *buf,
+ size_t len)
+{
+ unsigned int val;
+ int rc;
+
+ rc = kstrtouint(buf, 0, &val);
+ if (rc)
+ return rc;
+
+ guard(rwsem_write)(&tsm_rwsem);
+ if (tsm_report.desc.privlevel == val)
+ return len;
+ tsm_report.desc.privlevel = val;
+ tsm_report.write_generation++;
+
+ return len;
+}
+
+static ssize_t privlevel_show(struct device *dev, struct device_attribute *attr,
+ char *buf)
+{
+ return sysfs_emit(buf, "%u\n", tsm_report.desc.privlevel);
+}
+
+static DEVICE_ATTR_RW(privlevel);
+
+static ssize_t format_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t len)
+{
+ enum tsm_format format;
+
+ if (sysfs_streq(buf, "default"))
+ format = TSM_FORMAT_DEFAULT;
+ else if (sysfs_streq(buf, "extended"))
+ format = TSM_FORMAT_EXTENDED;
+ else
+ return -EINVAL;
+
+ guard(rwsem_write)(&tsm_rwsem);
+ if (tsm_report.desc.outblob_format == format)
+ return len;
+ tsm_report.desc.outblob_format = format;
+ tsm_report.write_generation++;
+
+ return len;
+}
+
+static ssize_t format_show(struct device *dev, struct device_attribute *attr,
+ char *buf)
+{
+ if (tsm_report.desc.outblob_format == TSM_FORMAT_DEFAULT)
+ return sysfs_emit(buf, "default\n");
+ return sysfs_emit(buf, "extended\n");
+}
+
+static DEVICE_ATTR_RW(format);
+
+static struct attribute *tsm_extra_attributes[] = {
+ &dev_attr_format.attr,
+ &dev_attr_privlevel.attr,
+ NULL,
+};
+
+struct attribute_group tsm_extra_attribute_group = {
+ .attrs = tsm_extra_attributes,
+};
+EXPORT_SYMBOL_GPL(tsm_extra_attribute_group);
+
+/*
+ * Input is a small hex blob, rather than a writable binary attribute, so that
+ * it is conveyed atomically.
+ */
+static ssize_t inhex_store(struct device *dev, struct device_attribute *attr,
+ const char *buf, size_t len)
+{
+ u8 inblob[TSM_INBLOB_MAX];
+ size_t inblob_len;
+ int rc;
+
+ inblob_len = len;
+ if (buf[len - 1] == '\n')
+ inblob_len--;
+ if (inblob_len & 1)
+ return -EINVAL;
+ inblob_len /= 2;
+ if (inblob_len > TSM_INBLOB_MAX)
+ return -EINVAL;
+
+ rc = hex2bin(inblob, buf, inblob_len);
+ if (rc < 0)
+ return rc;
+
+ guard(rwsem_write)(&tsm_rwsem);
+ if (memcmp(tsm_report.desc.inblob, inblob, inblob_len) == 0)
+ return len;
+ memcpy(tsm_report.desc.inblob, inblob, inblob_len);
+ tsm_report.desc.inblob_len = inblob_len;
+ tsm_report.write_generation++;
+
+ return len;
+}
+
+static ssize_t inhex_show(struct device *dev, struct device_attribute *attr,
+ char *buf)
+{
+ char *end;
+
+ guard(rwsem_read)(&tsm_rwsem);
+ if (!tsm_report.desc.inblob_len)
+ return 0;
+ end = bin2hex(buf, tsm_report.desc.inblob, tsm_report.desc.inblob_len);
+ *end++ = '\n';
+ return end - buf;
+}
+static DEVICE_ATTR_RW(inhex);
+
+static ssize_t generation_show(struct device *dev,
+ struct device_attribute *attr, char *buf)
+{
+ guard(rwsem_read)(&tsm_rwsem);
+ return sysfs_emit(buf, "%lu\n", tsm_report.write_generation);
+}
+static DEVICE_ATTR_RO(generation);
+
+static struct attribute *tsm_attributes[] = {
+ &dev_attr_inhex.attr,
+ &dev_attr_generation.attr,
+ NULL,
+};
+
+static ssize_t outblob_read(struct file *f, struct kobject *kobj,
+ struct bin_attribute *bin_attr, char *buf,
+ loff_t offset, size_t count)
+{
+ guard(rwsem_read)(&tsm_rwsem);
+ if (!tsm_report.desc.inblob_len)
+ return -EINVAL;
+
+ if (!tsm_report.outblob ||
+ tsm_report.read_generation != tsm_report.write_generation) {
+ const struct tsm_ops *ops = provider.ops;
+ size_t outblob_len;
+ u8 *outblob;
+
+ kvfree(tsm_report.outblob);
+ outblob = ops->report_new(provider.dev->parent,
+ &tsm_report.desc, &outblob_len);
+ if (IS_ERR(outblob))
+ return PTR_ERR(outblob);
+ tsm_report.outblob_len = outblob_len;
+ tsm_report.outblob = outblob;
+ tsm_report.read_generation = tsm_report.write_generation;
+ }
+
+ return memory_read_from_buffer(buf, count, &offset,
+ tsm_report.outblob,
+ tsm_report.outblob_len);
+}
+static BIN_ATTR_RO(outblob, 0);
+
+static struct bin_attribute *tsm_bin_attributes[] = {
+ &bin_attr_outblob,
+ NULL,
+};
+
+struct attribute_group tsm_default_attribute_group = {
+ .bin_attrs = tsm_bin_attributes,
+ .attrs = tsm_attributes,
+};
+EXPORT_SYMBOL_GPL(tsm_default_attribute_group);
+
+static const struct attribute_group *tsm_default_attribute_groups[] = {
+ &tsm_default_attribute_group,
+ NULL,
+};
+
+int register_tsm(const struct tsm_ops *ops, struct device *parent,
+ const struct attribute_group **groups)
+{
+ const struct tsm_ops *conflict;
+ struct device *dev;
+ int rc;
+
+ if (!parent)
+ return -EINVAL;
+
+ if (!groups)
+ groups = tsm_default_attribute_groups;
+
+ guard(rwsem_write)(&tsm_rwsem);
+ conflict = provider.ops;
+ if (conflict) {
+ pr_err("\"%s\" ops already registered\n", conflict->name);
+ return rc;
+ }
+
+ dev = device_create_with_groups(tsm_class, parent, 0, NULL, groups,
+ "tsm0");
+ if (IS_ERR(dev))
+ return PTR_ERR(dev);
+
+ provider.ops = ops;
+ provider.dev = dev;
+ return 0;
+}
+EXPORT_SYMBOL_GPL(register_tsm);
+
+int unregister_tsm(const struct tsm_ops *ops)
+{
+ guard(rwsem_write)(&tsm_rwsem);
+ if (ops != provider.ops)
+ return -EBUSY;
+ provider.ops = NULL;
+ device_unregister(provider.dev);
+ provider.dev = NULL;
+ kvfree(tsm_report.outblob);
+ tsm_report.outblob = NULL;
+ return 0;
+}
+EXPORT_SYMBOL_GPL(unregister_tsm);
+
+static int __init tsm_init(void)
+{
+ tsm_class = class_create("tsm");
+ return PTR_ERR_OR_ZERO(tsm_class);
+}
+module_init(tsm_init);
+
+static void __exit tsm_exit(void)
+{
+ class_destroy(tsm_class);
+}
+module_exit(tsm_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_DESCRIPTION("Provide Trusted Security Module attestation reports via sysfs");
diff --git a/include/linux/tsm.h b/include/linux/tsm.h
new file mode 100644
index 000000000000..6dc2f07543b8
--- /dev/null
+++ b/include/linux/tsm.h
@@ -0,0 +1,45 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef __TSM_H
+#define __TSM_H
+
+#include <linux/types.h>
+#include <linux/device.h>
+
+#define TSM_INBLOB_MAX 64
+
+enum tsm_format {
+ TSM_FORMAT_DEFAULT,
+ TSM_FORMAT_EXTENDED,
+};
+
+/**
+ * struct tsm_desc - option descriptor for generating tsm report blobs
+ * @privlevel: optional privilege level to associate with @outblob
+ * @inblob_len: sizeof @inblob
+ * @inblob: arbitrary input data
+ * @outblob_format: for TSMs with an "extended" format
+ */
+struct tsm_desc {
+ unsigned int privlevel;
+ size_t inblob_len;
+ u8 inblob[TSM_INBLOB_MAX];
+ enum tsm_format outblob_format;
+};
+
+/*
+ * arch specific ops, only one is expected to be registered at a time
+ * i.e. only one of SEV, TDX, COVE, etc.
+ */
+struct tsm_ops {
+ const char *name;
+ u8 *(*report_new)(struct device *dev, const struct tsm_desc *desc,
+ size_t *outblob_len);
+};
+
+extern struct attribute_group tsm_default_attribute_group;
+extern struct attribute_group tsm_extra_attribute_group;
+
+int register_tsm(const struct tsm_ops *ops, struct device *parent,
+ const struct attribute_group **groups);
+int unregister_tsm(const struct tsm_ops *ops);
+#endif /* __TSM_H */
In preparation for using the TSM key facility to convey attestation blobs
to userspace, add an argument to flag whether @arg is a user buffer or a
kernel buffer.
While TSM keys is meant to replace existing confidenital computing
ioctl() implementations for attestation report retrieval the old ioctl()
path needs to stick around for a deprecation period.
No behavior change intended, just introduce the copy wrappers and @type
argument.
Note that these wrappers are similar to copy_{to,from}_sockptr(). If
this approach moves forward that concept is something that can be
generalized into a helper with a generic name.
Cc: Borislav Petkov <[email protected]>
Cc: Tom Lendacky <[email protected]>
Cc: Dionna Glaze <[email protected]>
Cc: Brijesh Singh <[email protected]>
Signed-off-by: Dan Williams <[email protected]>
---
drivers/virt/coco/sev-guest/sev-guest.c | 48 ++++++++++++++++++++++++-------
1 file changed, 37 insertions(+), 11 deletions(-)
diff --git a/drivers/virt/coco/sev-guest/sev-guest.c b/drivers/virt/coco/sev-guest/sev-guest.c
index 97dbe715e96a..f48c4764a7a2 100644
--- a/drivers/virt/coco/sev-guest/sev-guest.c
+++ b/drivers/virt/coco/sev-guest/sev-guest.c
@@ -470,7 +470,32 @@ static int handle_guest_request(struct snp_guest_dev *snp_dev, u64 exit_code,
return 0;
}
-static int get_report(struct snp_guest_dev *snp_dev, struct snp_guest_request_ioctl *arg)
+enum snp_arg_type {
+ SNP_UARG,
+ SNP_KARG,
+};
+
+static unsigned long copy_from(void *to, unsigned long from, unsigned long n,
+ enum snp_arg_type type)
+{
+ if (type == SNP_UARG)
+ return copy_from_user(to, (void __user *)from, n);
+ memcpy(to, (void *)from, n);
+ return 0;
+}
+
+static unsigned long copy_to(unsigned long to, const void *from,
+ unsigned long n, enum snp_arg_type type)
+{
+ if (type == SNP_UARG)
+ return copy_to_user((void __user *)to, from, n);
+ memcpy((void *)to, from, n);
+ return 0;
+}
+
+static int get_report(struct snp_guest_dev *snp_dev,
+ struct snp_guest_request_ioctl *arg,
+ enum snp_arg_type type)
{
struct snp_guest_crypto *crypto = snp_dev->crypto;
struct snp_report_resp *resp;
@@ -482,7 +507,7 @@ static int get_report(struct snp_guest_dev *snp_dev, struct snp_guest_request_io
if (!arg->req_data || !arg->resp_data)
return -EINVAL;
- if (copy_from_user(&req, (void __user *)arg->req_data, sizeof(req)))
+ if (copy_from(&req, arg->req_data, sizeof(req), type))
return -EFAULT;
/*
@@ -501,7 +526,7 @@ static int get_report(struct snp_guest_dev *snp_dev, struct snp_guest_request_io
if (rc)
goto e_free;
- if (copy_to_user((void __user *)arg->resp_data, resp, sizeof(*resp)))
+ if (copy_to(arg->resp_data, resp, sizeof(*resp), type))
rc = -EFAULT;
e_free:
@@ -550,7 +575,9 @@ static int get_derived_key(struct snp_guest_dev *snp_dev, struct snp_guest_reque
return rc;
}
-static int get_ext_report(struct snp_guest_dev *snp_dev, struct snp_guest_request_ioctl *arg)
+static int get_ext_report(struct snp_guest_dev *snp_dev,
+ struct snp_guest_request_ioctl *arg,
+ enum snp_arg_type type)
{
struct snp_guest_crypto *crypto = snp_dev->crypto;
struct snp_ext_report_req req;
@@ -562,7 +589,7 @@ static int get_ext_report(struct snp_guest_dev *snp_dev, struct snp_guest_reques
if (!arg->req_data || !arg->resp_data)
return -EINVAL;
- if (copy_from_user(&req, (void __user *)arg->req_data, sizeof(req)))
+ if (copy_from(&req, arg->req_data, sizeof(req), type))
return -EFAULT;
/* userspace does not want certificate data */
@@ -611,14 +638,13 @@ static int get_ext_report(struct snp_guest_dev *snp_dev, struct snp_guest_reques
if (ret)
goto e_free;
- if (npages &&
- copy_to_user((void __user *)req.certs_address, snp_dev->certs_data,
- req.certs_len)) {
+ if (npages && copy_to(req.certs_address, snp_dev->certs_data,
+ req.certs_len, type)) {
ret = -EFAULT;
goto e_free;
}
- if (copy_to_user((void __user *)arg->resp_data, resp, sizeof(*resp)))
+ if (copy_to(arg->resp_data, resp, sizeof(*resp), type))
ret = -EFAULT;
e_free:
@@ -653,13 +679,13 @@ static long snp_guest_ioctl(struct file *file, unsigned int ioctl, unsigned long
switch (ioctl) {
case SNP_GET_REPORT:
- ret = get_report(snp_dev, &input);
+ ret = get_report(snp_dev, &input, SNP_UARG);
break;
case SNP_GET_DERIVED_KEY:
ret = get_derived_key(snp_dev, &input);
break;
case SNP_GET_EXT_REPORT:
- ret = get_ext_report(snp_dev, &input);
+ ret = get_ext_report(snp_dev, &input, SNP_UARG);
break;
default:
break;
On 8/14/2023 9:43 AM, Dan Williams wrote:
> Changes since v1:
> - Switch from Keyring to sysfs (James)
>
> An attestation report is signed evidence of how a Trusted Virtual
> Machine (TVM) was launched and its current state. A verifying party uses
> the report to make judgements of the confidentiality and integrity of
> that execution environment. Upon successful attestation the verifying
> party may, for example, proceed to deploy secrets to the TVM to carry
> out a workload. Multiple confidential computing platforms share this
> similar flow.
>
> The approach of adding adding new char devs and new ioctls, for what
> amounts to the same logical functionality with minor formatting
> differences across vendors [1], is untenable. Common concepts and the
> community benefit from common infrastructure.
>
> Use sysfs for this facility for maintainability compared to ioctl(). The
> expectation is that this interface is a boot time, configure once, get
> report, and done flow. I.e. not something that receives ongoing
> transactions at runtime. However, runtime retrieval is not precluded and
> a mechanism to detect potential configuration conflicts from
> multiple-threads using this interface is included.
>
I wanted to speak up to say that this does not align with the needs we have
in the Confidential Containers project. We want to be able to perform attestation
not just once during boot but during the lifecycle of the confidential VM. We
may need to fetch a fresh attestation report from a trusted agent but also from
arbitrary applications running in containers.
The trusted agent might need attestation when launching a new container from an
encrypted container image or when a new secret is being added to the VM - both
of these events may happen at any time (also when containerized applications
are already executing).
Container applications have their own uses for attestation, such as when they need
to fetch keys required to decrypt payloads. We also have things like performing
attestation when establishing a TLS or ssh connection to provide an attested e2e
encrypted communication channel.
I don't think sysfs is suitable for such concurrent transactions. Also if you think
about exposing the sysfs interface to an application in a container, this requires
bind mounting rw part of the sysfs tree into the mount namespace - not ideal.
Jeremi
> The keyring@ list is dropped on this posting since a new key-type is no
> longer being pursued.
>
> Link: http://lore.kernel.org/r/[email protected]
>
> ---
>
> Dan Williams (5):
> virt: coco: Add a coco/Makefile and coco/Kconfig
> tsm: Introduce a shared ABI for attestation reports
> virt: sevguest: Prep for kernel internal {get,get_ext}_report()
> mm/slab: Add __free() support for kvfree
> virt: sevguest: Add TSM_REPORTS support for SNP_{GET,GET_EXT}_REPORT
>
>
> Documentation/ABI/testing/sysfs-class-tsm | 47 +++++
> MAINTAINERS | 8 +
> drivers/virt/Kconfig | 6 -
> drivers/virt/Makefile | 4
> drivers/virt/coco/Kconfig | 13 +
> drivers/virt/coco/Makefile | 8 +
> drivers/virt/coco/sev-guest/Kconfig | 1
> drivers/virt/coco/sev-guest/sev-guest.c | 129 ++++++++++++-
> drivers/virt/coco/tdx-guest/Kconfig | 1
> drivers/virt/coco/tsm.c | 290 +++++++++++++++++++++++++++++
> include/linux/slab.h | 2
> include/linux/tsm.h | 45 +++++
> 12 files changed, 535 insertions(+), 19 deletions(-)
> create mode 100644 Documentation/ABI/testing/sysfs-class-tsm
> create mode 100644 drivers/virt/coco/Kconfig
> create mode 100644 drivers/virt/coco/Makefile
> create mode 100644 drivers/virt/coco/tsm.c
> create mode 100644 include/linux/tsm.h
>
> base-commit: 06c2afb862f9da8dc5efa4b6076a0e48c3fbaaa5
On 8/14/2023 9:43 AM, Dan Williams wrote:
> One of the common operations of a TSM (Trusted Security Module) is to
> provide a way for a TVM (confidential computing guest execution
> environment) to take a measurement of its launch state, sign it and
> submit it to a verifying party. Upon successful attestation that
> verifies the integrity of the TVM additional secrets may be deployed.
> The concept is common across TSMs, but the implementations are
> unfortunately vendor specific. While the industry grapples with a common
> definition of this attestation format [1], Linux need not make this
> problem worse by defining a new ABI per TSM that wants to perform a
> similar operation. The current momentum has been to invent new ioctl-ABI
> per TSM per function which at best is an abdication of the kernel's
> responsibility to make common infrastructure concepts share common ABI.
>
> The proposal, targeted to conceptually work with TDX, SEV, COVE if not
> more, is to define a sysfs interface to retrieve the TSM-specific blob.
>
> echo $hex_encoded_userdata_plus_nonce > /sys/class/tsm/tsm0/inhex
> hexdump /sys/class/tsm/tsm0/outblob
>
> This approach later allows for the standardization of the attestation
> blob format without needing to change the Linux ABI. Until then, the
> format of 'outblob' is determined by the parent device for 'tsm0'.
>
> The expectation is that this is a boot time exchange that need not be
> regenerated, making it amenable to a sysfs interface. In case userspace
> does try to generate multiple attestation reports it includes conflict
> detection so userspace can be sure no other thread changed the
> parameters from its last configuration step to the blob retrieval.
>
> TSM specific options are encoded as 'extra' attributes on the TSM device
> with the expectation that vendors reuse the same options for similar
> concepts. The current options are defined by SEV-SNP's need for a
> 'privilege level' concept (VMPL), and the option to retrieve a
> certificate chain in addition to the attestation report ("extended"
> format).
>
> Link: http://lore.kernel.org/r/[email protected] [1]
> Cc: Kuppuswamy Sathyanarayanan <[email protected]>
> Cc: Dionna Amalie Glaze <[email protected]>
> Cc: James Bottomley <[email protected]>
> Cc: Peter Gonda <[email protected]>
> Cc: Greg Kroah-Hartman <[email protected]>
> Cc: Samuel Ortiz <[email protected]>
> Signed-off-by: Dan Williams <[email protected]>
> ---
> Documentation/ABI/testing/sysfs-class-tsm | 47 +++++
> MAINTAINERS | 8 +
> drivers/virt/coco/Kconfig | 4
> drivers/virt/coco/Makefile | 1
> drivers/virt/coco/tdx-guest/Kconfig | 1
> drivers/virt/coco/tsm.c | 290 +++++++++++++++++++++++++++++
> include/linux/tsm.h | 45 +++++
> 7 files changed, 396 insertions(+)
> create mode 100644 Documentation/ABI/testing/sysfs-class-tsm
> create mode 100644 drivers/virt/coco/tsm.c
> create mode 100644 include/linux/tsm.h
>
> diff --git a/Documentation/ABI/testing/sysfs-class-tsm b/Documentation/ABI/testing/sysfs-class-tsm
> new file mode 100644
> index 000000000000..37017bde626d
> --- /dev/null
> +++ b/Documentation/ABI/testing/sysfs-class-tsm
> @@ -0,0 +1,47 @@
> +What: /sys/class/tsm/tsm0/inhex
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RW) Hex encoded userdata to be included in the attestation
> + report. For replay protection this should include a nonce, but
> + the kernel does not place any restrictions on the content.
> +
> +What: /sys/class/tsm/tsm0/outblob
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RO) Binary attestation report generated from @inhex translated
> + to binary and any options. The format of the report is vendor
> + specific and determined by the parent device of 'tsm0'.
> +
> +What: /sys/class/tsm/tsm0/generation
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RO) The value in this attribute increments each time @inhex or
> + any option is written. Userspace can detect conflicts by
> + checking generation before writing to any attribute and making
> + sure the number of writes matches expectations after reading
> + @outblob.
> +
> +What: /sys/class/tsm/tsm0/privlevel
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RW) If a TSM implementation supports the concept of attestation
> + reports for TVMs running at different privilege levels, like
> + SEV-SNP "VMPL", specify the privilege level via this attribute.
> +
> +What: /sys/class/tsm/tsm0/format
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RW) If a TSM implementation supports the concept of attestation
> + reports with "extended" contents, like SEV-SNP extended reports
> + with certificate chains, specify "extended" vs "default" via
> + this attribute.
> diff --git a/MAINTAINERS b/MAINTAINERS
> index 3be1bdfe8ecc..97f74d344c8a 100644
> --- a/MAINTAINERS
> +++ b/MAINTAINERS
> @@ -21625,6 +21625,14 @@ W: https://github.com/srcres258/linux-doc
> T: git git://github.com/srcres258/linux-doc.git doc-zh-tw
> F: Documentation/translations/zh_TW/
>
> +TRUSTED SECURITY MODULE (TSM) ATTESTATION REPORTS
> +M: Dan Williams <[email protected]>
> +L: [email protected]
> +S: Maintained
> +F: Documentation/ABI/testing/sysfs-class-tsm
> +F: drivers/virt/coco/tsm.c
> +F: include/linux/tsm.h
> +
> TTY LAYER
> M: Greg Kroah-Hartman <[email protected]>
> M: Jiri Slaby <[email protected]>
> diff --git a/drivers/virt/coco/Kconfig b/drivers/virt/coco/Kconfig
> index fc5c64f04c4a..d92f07019f38 100644
> --- a/drivers/virt/coco/Kconfig
> +++ b/drivers/virt/coco/Kconfig
> @@ -2,6 +2,10 @@
> #
> # Confidential computing related collateral
> #
> +
> +config TSM_REPORTS
> + tristate
> +
> source "drivers/virt/coco/efi_secret/Kconfig"
>
> source "drivers/virt/coco/sev-guest/Kconfig"
> diff --git a/drivers/virt/coco/Makefile b/drivers/virt/coco/Makefile
> index 55302ef719ad..18c1aba5edb7 100644
> --- a/drivers/virt/coco/Makefile
> +++ b/drivers/virt/coco/Makefile
> @@ -2,6 +2,7 @@
> #
> # Confidential computing related collateral
> #
> +obj-$(CONFIG_TSM_REPORTS) += tsm.o
> obj-$(CONFIG_EFI_SECRET) += efi_secret/
> obj-$(CONFIG_SEV_GUEST) += sev-guest/
> obj-$(CONFIG_INTEL_TDX_GUEST) += tdx-guest/
> diff --git a/drivers/virt/coco/tdx-guest/Kconfig b/drivers/virt/coco/tdx-guest/Kconfig
> index 14246fc2fb02..22dd59e19431 100644
> --- a/drivers/virt/coco/tdx-guest/Kconfig
> +++ b/drivers/virt/coco/tdx-guest/Kconfig
> @@ -1,6 +1,7 @@
> config TDX_GUEST_DRIVER
> tristate "TDX Guest driver"
> depends on INTEL_TDX_GUEST
> + select TSM_REPORTS
> help
> The driver provides userspace interface to communicate with
> the TDX module to request the TDX guest details like attestation
> diff --git a/drivers/virt/coco/tsm.c b/drivers/virt/coco/tsm.c
> new file mode 100644
> index 000000000000..1bf2ee82eb94
> --- /dev/null
> +++ b/drivers/virt/coco/tsm.c
> @@ -0,0 +1,290 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/* Copyright(c) 2023 Intel Corporation. All rights reserved. */
> +
> +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
> +
> +#include <linux/tsm.h>
> +#include <linux/err.h>
> +#include <linux/slab.h>
> +#include <linux/rwsem.h>
> +#include <linux/device.h>
> +#include <linux/string.h>
> +#include <linux/module.h>
> +#include <linux/cleanup.h>
> +
> +struct class *tsm_class;
> +static struct tsm_provider {
> + const struct tsm_ops *ops;
> + struct device *dev;
> +} provider;
> +static DECLARE_RWSEM(tsm_rwsem);
> +
> +/**
> + * DOC: Trusted Security Module (TSM) Attestation Report Interface
> + *
> + * The TSM report interface is a common provider of blobs that facilitate
> + * attestation of a TVM (confidential computing guest) by an attestation
> + * service. A TSM report combines a user-defined blob (likely a public-key with
> + * a nonce for a key-exchange protocol) with a signed attestation report. That
> + * combined blob is then used to obtain secrets provided by an agent that can
> + * validate the attestation report. The expectation is that this interface is
> + * invoked infrequently, likely only once at TVM boot time.
> + *
> + * The attestation report format is TSM provider specific, when / if a standard
> + * materializes that can be published instead of the vendor layout.
> + */
> +
> +/**
> + * struct tsm_report - track state of report generation relative to options
> + * @desc: report generation options / cached report state
> + * @outblob: generated evidence to provider to the attestation agent
> + * @outblob_len: sizeof(outblob)
> + * @write_generation: conflict detection, and report regeneration tracking
> + * @read_generation: cached report invalidation tracking
> + */
> +struct tsm_report {
> + struct tsm_desc desc;
> + size_t outblob_len;
> + u8 *outblob;
> + unsigned long write_generation;
> + unsigned long read_generation;
> +} tsm_report;
> +
> +static ssize_t privlevel_store(struct device *dev,
> + struct device_attribute *attr, const char *buf,
> + size_t len)
> +{
> + unsigned int val;
> + int rc;
> +
> + rc = kstrtouint(buf, 0, &val);
> + if (rc)
> + return rc;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + if (tsm_report.desc.privlevel == val)
> + return len;
> + tsm_report.desc.privlevel = val;
> + tsm_report.write_generation++;
> +
> + return len;
> +}
> +
> +static ssize_t privlevel_show(struct device *dev, struct device_attribute *attr,
> + char *buf)
> +{
> + return sysfs_emit(buf, "%u\n", tsm_report.desc.privlevel);
> +}
> +
> +static DEVICE_ATTR_RW(privlevel);
> +
> +static ssize_t format_store(struct device *dev, struct device_attribute *attr,
> + const char *buf, size_t len)
> +{
> + enum tsm_format format;
> +
> + if (sysfs_streq(buf, "default"))
> + format = TSM_FORMAT_DEFAULT;
> + else if (sysfs_streq(buf, "extended"))
> + format = TSM_FORMAT_EXTENDED;
> + else
> + return -EINVAL;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + if (tsm_report.desc.outblob_format == format)
> + return len;
> + tsm_report.desc.outblob_format = format;
> + tsm_report.write_generation++;
> +
> + return len;
> +}
> +
> +static ssize_t format_show(struct device *dev, struct device_attribute *attr,
> + char *buf)
> +{
> + if (tsm_report.desc.outblob_format == TSM_FORMAT_DEFAULT)
> + return sysfs_emit(buf, "default\n");
> + return sysfs_emit(buf, "extended\n");
> +}
> +
> +static DEVICE_ATTR_RW(format);
> +
> +static struct attribute *tsm_extra_attributes[] = {
> + &dev_attr_format.attr,
> + &dev_attr_privlevel.attr,
> + NULL,
> +};
> +
> +struct attribute_group tsm_extra_attribute_group = {
> + .attrs = tsm_extra_attributes,
> +};
> +EXPORT_SYMBOL_GPL(tsm_extra_attribute_group);
> +
> +/*
> + * Input is a small hex blob, rather than a writable binary attribute, so that
> + * it is conveyed atomically.
> + */
> +static ssize_t inhex_store(struct device *dev, struct device_attribute *attr,
> + const char *buf, size_t len)
> +{
> + u8 inblob[TSM_INBLOB_MAX];
> + size_t inblob_len;
> + int rc;
> +
> + inblob_len = len;
> + if (buf[len - 1] == '\n')
> + inblob_len--;
> + if (inblob_len & 1)
> + return -EINVAL;
> + inblob_len /= 2;
> + if (inblob_len > TSM_INBLOB_MAX)
> + return -EINVAL;
> +
> + rc = hex2bin(inblob, buf, inblob_len);
> + if (rc < 0)
> + return rc;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + if (memcmp(tsm_report.desc.inblob, inblob, inblob_len) == 0)
> + return len;
> + memcpy(tsm_report.desc.inblob, inblob, inblob_len);
> + tsm_report.desc.inblob_len = inblob_len;
> + tsm_report.write_generation++;
> +
> + return len;
> +}
> +
> +static ssize_t inhex_show(struct device *dev, struct device_attribute *attr,
> + char *buf)
> +{
> + char *end;
> +
> + guard(rwsem_read)(&tsm_rwsem);
> + if (!tsm_report.desc.inblob_len)
> + return 0;
> + end = bin2hex(buf, tsm_report.desc.inblob, tsm_report.desc.inblob_len);
> + *end++ = '\n';
> + return end - buf;
> +}
> +static DEVICE_ATTR_RW(inhex);
> +
> +static ssize_t generation_show(struct device *dev,
> + struct device_attribute *attr, char *buf)
> +{
> + guard(rwsem_read)(&tsm_rwsem);
> + return sysfs_emit(buf, "%lu\n", tsm_report.write_generation);
> +}
> +static DEVICE_ATTR_RO(generation);
> +
> +static struct attribute *tsm_attributes[] = {
> + &dev_attr_inhex.attr,
> + &dev_attr_generation.attr,
> + NULL,
> +};
> +
> +static ssize_t outblob_read(struct file *f, struct kobject *kobj,
> + struct bin_attribute *bin_attr, char *buf,
> + loff_t offset, size_t count)
> +{
> + guard(rwsem_read)(&tsm_rwsem);
This is unfortunate but it would need to be a rwsem_write otherwise two
processes can race to reach the kvfree and both call report_new at the
same time (unlikely as it may be).
Jeremi
> + if (!tsm_report.desc.inblob_len)
> + return -EINVAL;
> +
> + if (!tsm_report.outblob ||
> + tsm_report.read_generation != tsm_report.write_generation) {
> + const struct tsm_ops *ops = provider.ops;
> + size_t outblob_len;
> + u8 *outblob;
> +
> + kvfree(tsm_report.outblob);
> + outblob = ops->report_new(provider.dev->parent,
> + &tsm_report.desc, &outblob_len);
> + if (IS_ERR(outblob))
> + return PTR_ERR(outblob);
> + tsm_report.outblob_len = outblob_len;
> + tsm_report.outblob = outblob;
> + tsm_report.read_generation = tsm_report.write_generation;
> + }
> +
> + return memory_read_from_buffer(buf, count, &offset,
> + tsm_report.outblob,
> + tsm_report.outblob_len);
> +}
> +static BIN_ATTR_RO(outblob, 0);
> +
> +static struct bin_attribute *tsm_bin_attributes[] = {
> + &bin_attr_outblob,
> + NULL,
> +};
> +
> +struct attribute_group tsm_default_attribute_group = {
> + .bin_attrs = tsm_bin_attributes,
> + .attrs = tsm_attributes,
> +};
> +EXPORT_SYMBOL_GPL(tsm_default_attribute_group);
> +
> +static const struct attribute_group *tsm_default_attribute_groups[] = {
> + &tsm_default_attribute_group,
> + NULL,
> +};
> +
> +int register_tsm(const struct tsm_ops *ops, struct device *parent,
> + const struct attribute_group **groups)
> +{
> + const struct tsm_ops *conflict;
> + struct device *dev;
> + int rc;
> +
> + if (!parent)
> + return -EINVAL;
> +
> + if (!groups)
> + groups = tsm_default_attribute_groups;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + conflict = provider.ops;
> + if (conflict) {
> + pr_err("\"%s\" ops already registered\n", conflict->name);
> + return rc;
> + }
> +
> + dev = device_create_with_groups(tsm_class, parent, 0, NULL, groups,
> + "tsm0");
> + if (IS_ERR(dev))
> + return PTR_ERR(dev);
> +
> + provider.ops = ops;
> + provider.dev = dev;
> + return 0;
> +}
> +EXPORT_SYMBOL_GPL(register_tsm);
> +
> +int unregister_tsm(const struct tsm_ops *ops)
> +{
> + guard(rwsem_write)(&tsm_rwsem);
> + if (ops != provider.ops)
> + return -EBUSY;
> + provider.ops = NULL;
> + device_unregister(provider.dev);
> + provider.dev = NULL;
> + kvfree(tsm_report.outblob);
> + tsm_report.outblob = NULL;
> + return 0;
> +}
> +EXPORT_SYMBOL_GPL(unregister_tsm);
> +
> +static int __init tsm_init(void)
> +{
> + tsm_class = class_create("tsm");
> + return PTR_ERR_OR_ZERO(tsm_class);
> +}
> +module_init(tsm_init);
> +
> +static void __exit tsm_exit(void)
> +{
> + class_destroy(tsm_class);
> +}
> +module_exit(tsm_exit);
> +
> +MODULE_LICENSE("GPL");
> +MODULE_DESCRIPTION("Provide Trusted Security Module attestation reports via sysfs");
> diff --git a/include/linux/tsm.h b/include/linux/tsm.h
> new file mode 100644
> index 000000000000..6dc2f07543b8
> --- /dev/null
> +++ b/include/linux/tsm.h
> @@ -0,0 +1,45 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef __TSM_H
> +#define __TSM_H
> +
> +#include <linux/types.h>
> +#include <linux/device.h>
> +
> +#define TSM_INBLOB_MAX 64
> +
> +enum tsm_format {
> + TSM_FORMAT_DEFAULT,
> + TSM_FORMAT_EXTENDED,
> +};
> +
> +/**
> + * struct tsm_desc - option descriptor for generating tsm report blobs
> + * @privlevel: optional privilege level to associate with @outblob
> + * @inblob_len: sizeof @inblob
> + * @inblob: arbitrary input data
> + * @outblob_format: for TSMs with an "extended" format
> + */
> +struct tsm_desc {
> + unsigned int privlevel;
> + size_t inblob_len;
> + u8 inblob[TSM_INBLOB_MAX];
> + enum tsm_format outblob_format;
> +};
> +
> +/*
> + * arch specific ops, only one is expected to be registered at a time
> + * i.e. only one of SEV, TDX, COVE, etc.
> + */
> +struct tsm_ops {
> + const char *name;
> + u8 *(*report_new)(struct device *dev, const struct tsm_desc *desc,
> + size_t *outblob_len);
> +};
> +
> +extern struct attribute_group tsm_default_attribute_group;
> +extern struct attribute_group tsm_extra_attribute_group;
> +
> +int register_tsm(const struct tsm_ops *ops, struct device *parent,
> + const struct attribute_group **groups);
> +int unregister_tsm(const struct tsm_ops *ops);
> +#endif /* __TSM_H */
>
On Mon, Aug 14, 2023 at 12:43:21AM -0700, Dan Williams wrote:
> One of the common operations of a TSM (Trusted Security Module) is to
> provide a way for a TVM (confidential computing guest execution
> environment) to take a measurement of its launch state, sign it and
> submit it to a verifying party. Upon successful attestation that
> verifies the integrity of the TVM additional secrets may be deployed.
> The concept is common across TSMs, but the implementations are
> unfortunately vendor specific. While the industry grapples with a common
> definition of this attestation format [1], Linux need not make this
> problem worse by defining a new ABI per TSM that wants to perform a
> similar operation. The current momentum has been to invent new ioctl-ABI
> per TSM per function which at best is an abdication of the kernel's
> responsibility to make common infrastructure concepts share common ABI.
>
> The proposal, targeted to conceptually work with TDX, SEV, COVE if not
> more, is to define a sysfs interface to retrieve the TSM-specific blob.
>
> echo $hex_encoded_userdata_plus_nonce > /sys/class/tsm/tsm0/inhex
> hexdump /sys/class/tsm/tsm0/outblob
Why is one way a hex-encode file, that the kernel has to parse, and the
other not? Binary sysfs files should be "pass through" if at all
possible, why not make them both binary and not mess with hex at all?
That keeps the kernel simpler, and if userspace wants the hex format,
they can provide it much easier (with less potential parsing errors).
>
> This approach later allows for the standardization of the attestation
> blob format without needing to change the Linux ABI. Until then, the
> format of 'outblob' is determined by the parent device for 'tsm0'.
>
> The expectation is that this is a boot time exchange that need not be
> regenerated, making it amenable to a sysfs interface. In case userspace
> does try to generate multiple attestation reports it includes conflict
> detection so userspace can be sure no other thread changed the
> parameters from its last configuration step to the blob retrieval.
>
> TSM specific options are encoded as 'extra' attributes on the TSM device
> with the expectation that vendors reuse the same options for similar
> concepts. The current options are defined by SEV-SNP's need for a
> 'privilege level' concept (VMPL), and the option to retrieve a
> certificate chain in addition to the attestation report ("extended"
> format).
>
> Link: http://lore.kernel.org/r/[email protected] [1]
> Cc: Kuppuswamy Sathyanarayanan <[email protected]>
> Cc: Dionna Amalie Glaze <[email protected]>
> Cc: James Bottomley <[email protected]>
> Cc: Peter Gonda <[email protected]>
> Cc: Greg Kroah-Hartman <[email protected]>
> Cc: Samuel Ortiz <[email protected]>
> Signed-off-by: Dan Williams <[email protected]>
> ---
> Documentation/ABI/testing/sysfs-class-tsm | 47 +++++
> MAINTAINERS | 8 +
> drivers/virt/coco/Kconfig | 4
> drivers/virt/coco/Makefile | 1
> drivers/virt/coco/tdx-guest/Kconfig | 1
> drivers/virt/coco/tsm.c | 290 +++++++++++++++++++++++++++++
> include/linux/tsm.h | 45 +++++
> 7 files changed, 396 insertions(+)
> create mode 100644 Documentation/ABI/testing/sysfs-class-tsm
> create mode 100644 drivers/virt/coco/tsm.c
> create mode 100644 include/linux/tsm.h
>
> diff --git a/Documentation/ABI/testing/sysfs-class-tsm b/Documentation/ABI/testing/sysfs-class-tsm
> new file mode 100644
> index 000000000000..37017bde626d
> --- /dev/null
> +++ b/Documentation/ABI/testing/sysfs-class-tsm
> @@ -0,0 +1,47 @@
> +What: /sys/class/tsm/tsm0/inhex
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RW) Hex encoded userdata to be included in the attestation
> + report. For replay protection this should include a nonce, but
> + the kernel does not place any restrictions on the content.
"inhex" and it's read/write? Naming is hard :(
> +
> +What: /sys/class/tsm/tsm0/outblob
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RO) Binary attestation report generated from @inhex translated
> + to binary and any options. The format of the report is vendor
> + specific and determined by the parent device of 'tsm0'.
> +
> +What: /sys/class/tsm/tsm0/generation
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RO) The value in this attribute increments each time @inhex or
> + any option is written. Userspace can detect conflicts by
> + checking generation before writing to any attribute and making
> + sure the number of writes matches expectations after reading
> + @outblob.
> +
> +What: /sys/class/tsm/tsm0/privlevel
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RW) If a TSM implementation supports the concept of attestation
> + reports for TVMs running at different privilege levels, like
> + SEV-SNP "VMPL", specify the privilege level via this attribute.
Where is the list of potential values for this file at?
> +
> +What: /sys/class/tsm/tsm0/format
> +Date: August, 2023
> +KernelVersion: v6.6
> +Contact: [email protected]
> +Description:
> + (RW) If a TSM implementation supports the concept of attestation
> + reports with "extended" contents, like SEV-SNP extended reports
> + with certificate chains, specify "extended" vs "default" via
> + this attribute.
> diff --git a/MAINTAINERS b/MAINTAINERS
> index 3be1bdfe8ecc..97f74d344c8a 100644
> --- a/MAINTAINERS
> +++ b/MAINTAINERS
> @@ -21625,6 +21625,14 @@ W: https://github.com/srcres258/linux-doc
> T: git git://github.com/srcres258/linux-doc.git doc-zh-tw
> F: Documentation/translations/zh_TW/
>
> +TRUSTED SECURITY MODULE (TSM) ATTESTATION REPORTS
> +M: Dan Williams <[email protected]>
> +L: [email protected]
> +S: Maintained
> +F: Documentation/ABI/testing/sysfs-class-tsm
> +F: drivers/virt/coco/tsm.c
> +F: include/linux/tsm.h
> +
> TTY LAYER
> M: Greg Kroah-Hartman <[email protected]>
> M: Jiri Slaby <[email protected]>
> diff --git a/drivers/virt/coco/Kconfig b/drivers/virt/coco/Kconfig
> index fc5c64f04c4a..d92f07019f38 100644
> --- a/drivers/virt/coco/Kconfig
> +++ b/drivers/virt/coco/Kconfig
> @@ -2,6 +2,10 @@
> #
> # Confidential computing related collateral
> #
> +
> +config TSM_REPORTS
> + tristate
> +
> source "drivers/virt/coco/efi_secret/Kconfig"
>
> source "drivers/virt/coco/sev-guest/Kconfig"
> diff --git a/drivers/virt/coco/Makefile b/drivers/virt/coco/Makefile
> index 55302ef719ad..18c1aba5edb7 100644
> --- a/drivers/virt/coco/Makefile
> +++ b/drivers/virt/coco/Makefile
> @@ -2,6 +2,7 @@
> #
> # Confidential computing related collateral
> #
> +obj-$(CONFIG_TSM_REPORTS) += tsm.o
> obj-$(CONFIG_EFI_SECRET) += efi_secret/
> obj-$(CONFIG_SEV_GUEST) += sev-guest/
> obj-$(CONFIG_INTEL_TDX_GUEST) += tdx-guest/
> diff --git a/drivers/virt/coco/tdx-guest/Kconfig b/drivers/virt/coco/tdx-guest/Kconfig
> index 14246fc2fb02..22dd59e19431 100644
> --- a/drivers/virt/coco/tdx-guest/Kconfig
> +++ b/drivers/virt/coco/tdx-guest/Kconfig
> @@ -1,6 +1,7 @@
> config TDX_GUEST_DRIVER
> tristate "TDX Guest driver"
> depends on INTEL_TDX_GUEST
> + select TSM_REPORTS
> help
> The driver provides userspace interface to communicate with
> the TDX module to request the TDX guest details like attestation
> diff --git a/drivers/virt/coco/tsm.c b/drivers/virt/coco/tsm.c
> new file mode 100644
> index 000000000000..1bf2ee82eb94
> --- /dev/null
> +++ b/drivers/virt/coco/tsm.c
> @@ -0,0 +1,290 @@
> +// SPDX-License-Identifier: GPL-2.0-only
> +/* Copyright(c) 2023 Intel Corporation. All rights reserved. */
> +
> +#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
> +
> +#include <linux/tsm.h>
> +#include <linux/err.h>
> +#include <linux/slab.h>
> +#include <linux/rwsem.h>
> +#include <linux/device.h>
> +#include <linux/string.h>
> +#include <linux/module.h>
> +#include <linux/cleanup.h>
> +
> +struct class *tsm_class;
Nit, we are moving away from using class_create(), please make this a
const static class and register it with the driver core instead. That
way we don't have to fix it up in the future when we finally catch up
with all of the existing class_create() calls we have left.
See the patches in this -rc cycle for a bunch of them already, with many
more on lkml for examples of how to convert this. Here's one example:
https://lore.kernel.org/r/[email protected]
> +static struct tsm_provider {
> + const struct tsm_ops *ops;
> + struct device *dev;
> +} provider;
> +static DECLARE_RWSEM(tsm_rwsem);
> +
> +/**
> + * DOC: Trusted Security Module (TSM) Attestation Report Interface
> + *
> + * The TSM report interface is a common provider of blobs that facilitate
> + * attestation of a TVM (confidential computing guest) by an attestation
> + * service. A TSM report combines a user-defined blob (likely a public-key with
> + * a nonce for a key-exchange protocol) with a signed attestation report. That
> + * combined blob is then used to obtain secrets provided by an agent that can
> + * validate the attestation report. The expectation is that this interface is
> + * invoked infrequently, likely only once at TVM boot time.
> + *
> + * The attestation report format is TSM provider specific, when / if a standard
> + * materializes that can be published instead of the vendor layout.
Published where?
> +/*
> + * Input is a small hex blob, rather than a writable binary attribute, so that
> + * it is conveyed atomically.
> + */
> +static ssize_t inhex_store(struct device *dev, struct device_attribute *attr,
> + const char *buf, size_t len)
> +{
> + u8 inblob[TSM_INBLOB_MAX];
Can this get too big for the stack?
> +static ssize_t inhex_show(struct device *dev, struct device_attribute *attr,
> + char *buf)
> +{
> + char *end;
> +
> + guard(rwsem_read)(&tsm_rwsem);
I like seeing the guard() usage, very nice :)
Overall, the sysfs api is ok, except for the hex values, which is easy
to change. The usage of sysfs is ok as well, no complaints from me
there.
thanks,
greg k-h
>
> switch (ioctl) {
> case SNP_GET_REPORT:
> - ret = get_report(snp_dev, &input);
> + ret = get_report(snp_dev, &input, SNP_UARG);
> break;
> case SNP_GET_DERIVED_KEY:
> ret = get_derived_key(snp_dev, &input);
> break;
Do we have an agreement around the continued existence of sev-guest
for supporting derived keys, is that similarly slated for the chopping
block, or is it left undecided?
It appears your choice to not include the uarg/karg extension here is
deliberate.
> case SNP_GET_EXT_REPORT:
> - ret = get_ext_report(snp_dev, &input);
> + ret = get_ext_report(snp_dev, &input, SNP_UARG);
> break;
> default:
> break;
>
Reviewed-by: Dionna Glaze <[email protected]>
--
-Dionna Glaze, PhD (she/her)
Jeremi Piotrowski wrote:
> On 8/14/2023 9:43 AM, Dan Williams wrote:
> > Changes since v1:
> > - Switch from Keyring to sysfs (James)
> >
> > An attestation report is signed evidence of how a Trusted Virtual
> > Machine (TVM) was launched and its current state. A verifying party uses
> > the report to make judgements of the confidentiality and integrity of
> > that execution environment. Upon successful attestation the verifying
> > party may, for example, proceed to deploy secrets to the TVM to carry
> > out a workload. Multiple confidential computing platforms share this
> > similar flow.
> >
> > The approach of adding adding new char devs and new ioctls, for what
> > amounts to the same logical functionality with minor formatting
> > differences across vendors [1], is untenable. Common concepts and the
> > community benefit from common infrastructure.
> >
> > Use sysfs for this facility for maintainability compared to ioctl(). The
> > expectation is that this interface is a boot time, configure once, get
> > report, and done flow. I.e. not something that receives ongoing
> > transactions at runtime. However, runtime retrieval is not precluded and
> > a mechanism to detect potential configuration conflicts from
> > multiple-threads using this interface is included.
> >
>
> I wanted to speak up to say that this does not align with the needs we have
> in the Confidential Containers project. We want to be able to perform attestation
> not just once during boot but during the lifecycle of the confidential VM. We
> may need to fetch a fresh attestation report from a trusted agent but also from
> arbitrary applications running in containers.
>
> The trusted agent might need attestation when launching a new container from an
> encrypted container image or when a new secret is being added to the VM - both
> of these events may happen at any time (also when containerized applications
> are already executing).
>
> Container applications have their own uses for attestation, such as when they need
> to fetch keys required to decrypt payloads. We also have things like performing
> attestation when establishing a TLS or ssh connection to provide an attested e2e
> encrypted communication channel.
...and you expect that the boot time attestation becomes invalidated
later at run time such that ongoing round trips to the TSM are needed? I
am looking at "Table 21. ATTESTATION_REPORT Structure" for example and
not seeing data there that changes from one request to the next. Runtime
validation likely looks more like the vTPM use case with PCRs. That will
leverage the existing / common TPM ABI.
> I don't think sysfs is suitable for such concurrent transactions. Also if you think
> about exposing the sysfs interface to an application in a container, this requires
> bind mounting rw part of the sysfs tree into the mount namespace - not ideal.
sysfs is not suitable for concurrent transactions. The container would
need to have an alternate path to request that the singleton owner of
the interface generate new reports, or use the boot time attestation to
derive per container communication sessions to the attestation agent.
Jeremi Piotrowski wrote:
> On 8/14/2023 9:43 AM, Dan Williams wrote:
> > One of the common operations of a TSM (Trusted Security Module) is to
> > provide a way for a TVM (confidential computing guest execution
> > environment) to take a measurement of its launch state, sign it and
> > submit it to a verifying party. Upon successful attestation that
> > verifies the integrity of the TVM additional secrets may be deployed.
> > The concept is common across TSMs, but the implementations are
> > unfortunately vendor specific. While the industry grapples with a common
> > definition of this attestation format [1], Linux need not make this
> > problem worse by defining a new ABI per TSM that wants to perform a
> > similar operation. The current momentum has been to invent new ioctl-ABI
> > per TSM per function which at best is an abdication of the kernel's
> > responsibility to make common infrastructure concepts share common ABI.
> >
> > The proposal, targeted to conceptually work with TDX, SEV, COVE if not
> > more, is to define a sysfs interface to retrieve the TSM-specific blob.
> >
> > echo $hex_encoded_userdata_plus_nonce > /sys/class/tsm/tsm0/inhex
> > hexdump /sys/class/tsm/tsm0/outblob
> >
> > This approach later allows for the standardization of the attestation
> > blob format without needing to change the Linux ABI. Until then, the
> > format of 'outblob' is determined by the parent device for 'tsm0'.
> >
> > The expectation is that this is a boot time exchange that need not be
> > regenerated, making it amenable to a sysfs interface. In case userspace
> > does try to generate multiple attestation reports it includes conflict
> > detection so userspace can be sure no other thread changed the
> > parameters from its last configuration step to the blob retrieval.
> >
> > TSM specific options are encoded as 'extra' attributes on the TSM device
> > with the expectation that vendors reuse the same options for similar
> > concepts. The current options are defined by SEV-SNP's need for a
> > 'privilege level' concept (VMPL), and the option to retrieve a
> > certificate chain in addition to the attestation report ("extended"
> > format).
> >
> > Link: http://lore.kernel.org/r/[email protected] [1]
> > Cc: Kuppuswamy Sathyanarayanan <[email protected]>
> > Cc: Dionna Amalie Glaze <[email protected]>
> > Cc: James Bottomley <[email protected]>
> > Cc: Peter Gonda <[email protected]>
> > Cc: Greg Kroah-Hartman <[email protected]>
> > Cc: Samuel Ortiz <[email protected]>
> > Signed-off-by: Dan Williams <[email protected]>
> > ---
[..]
> > +static ssize_t outblob_read(struct file *f, struct kobject *kobj,
> > + struct bin_attribute *bin_attr, char *buf,
> > + loff_t offset, size_t count)
> > +{
> > + guard(rwsem_read)(&tsm_rwsem);
>
> This is unfortunate but it would need to be a rwsem_write otherwise two
> processes can race to reach the kvfree and both call report_new at the
> same time (unlikely as it may be).
Ugh, yup, good eye, will fix.
>
> Why do we need to be so prescriptive about "boot time" vs "runtime"
> attestations? A user may wish to attest to several requests as Jeremi
> notes. And why should users be forced into using a vTPM interface if
> their usecase doesn't require all the features and complexity of a
> vTPM? Some users may prefer less overall code within their Trusted
> Computer Base (TCB) and a TPM emulate is a significant code base.
>
I agree, and I was a bit too hasty to acquiesce to sysfs due to the
TPM argument that really only applies for SEV-SNP without a whole lot
of extra work for other backends (not to say SVSM isn't itself a whole
lot of extra work).
> It seems like you are just reading the SNP spec, if you read the TDX
> spec you'll see there are RTMRs which can be extended with new data.
> This leads to a different data in the attestation. Similar there are
> REMs in the ARM CCA spec.
>
I'll add a note here that measurement registers are extensible at any
point by ring0, and there should be an API for doing so, the way that
there is for /dev/tpmX.
It could be /dev/teemr or something to unify TDX, COVE, ARM CCA, and
potentially a measurement register protocol extension to SEV-SNP's
SVSM.
I'm not sure how Intel is going to propose abstracting TCG Canonical
Event Log measurements to reuse measurement-to-PCR<X> code points in
the kernel as measurement-to-MR<f(X)>, or whatnot, but each technology
should have that implementation option to extend their own measurement
registers (and event log, potentially).
I (and probably James) object with just saying the PCRs are going to
xyz-measurement-register for simulating that integrity part of a TPM
to get just the quote aspect and not the rest of TPM 2.0 to hide
everything behind the TPM abstraction. It doesn't follow the Tcg spec.
But I repeat myself. If we use any ioctl, we'll end up multiplexing
the input per-technology, and at that point we essentially have
manufacturer-specific devices much to Dan's dismay.
Sysfs will certainly not be okay for measurement register-only
technology, since there's no way to not use a hardware attestation to
securely track measurement changes past "the static boot" (PCRs 0-7).
I don't want to have to rely on enclave-like peer VMs that perform the
TPM behavior.
--
-Dionna Glaze, PhD (she/her)
On 8/14/23 02:43, Dan Williams wrote:
> In preparation for using the TSM key facility to convey attestation blobs
> to userspace, add an argument to flag whether @arg is a user buffer or a
> kernel buffer.
>
> While TSM keys is meant to replace existing confidenital computing
s/confidenital/confidential/
> ioctl() implementations for attestation report retrieval the old ioctl()
> path needs to stick around for a deprecation period.
>
> No behavior change intended, just introduce the copy wrappers and @type
> argument.
>
> Note that these wrappers are similar to copy_{to,from}_sockptr(). If
> this approach moves forward that concept is something that can be
> generalized into a helper with a generic name.
>
> Cc: Borislav Petkov <[email protected]>
> Cc: Tom Lendacky <[email protected]>
> Cc: Dionna Glaze <[email protected]>
> Cc: Brijesh Singh <[email protected]>
> Signed-off-by: Dan Williams <[email protected]>
> ---
> drivers/virt/coco/sev-guest/sev-guest.c | 48 ++++++++++++++++++++++++-------
> 1 file changed, 37 insertions(+), 11 deletions(-)
>
> diff --git a/drivers/virt/coco/sev-guest/sev-guest.c b/drivers/virt/coco/sev-guest/sev-guest.c
> index 97dbe715e96a..f48c4764a7a2 100644
> --- a/drivers/virt/coco/sev-guest/sev-guest.c
> +++ b/drivers/virt/coco/sev-guest/sev-guest.c
> @@ -470,7 +470,32 @@ static int handle_guest_request(struct snp_guest_dev *snp_dev, u64 exit_code,
> return 0;
> }
>
> -static int get_report(struct snp_guest_dev *snp_dev, struct snp_guest_request_ioctl *arg)
> +enum snp_arg_type {
> + SNP_UARG,
> + SNP_KARG,
> +};
> +
> +static unsigned long copy_from(void *to, unsigned long from, unsigned long n,
> + enum snp_arg_type type)
> +{
> + if (type == SNP_UARG)
> + return copy_from_user(to, (void __user *)from, n);
I'm a fan of blank lines to make reading functions easier. A blank line
here and below after the memcpy() would be nice.
Ditto in the copy_to() function.
> + memcpy(to, (void *)from, n);
> + return 0;
> +}
> +
> +static unsigned long copy_to(unsigned long to, const void *from,
> + unsigned long n, enum snp_arg_type type)
> +{
> + if (type == SNP_UARG)
> + return copy_to_user((void __user *)to, from, n);
> + memcpy((void *)to, from, n);
> + return 0;
> +}
> +
> +static int get_report(struct snp_guest_dev *snp_dev,
> + struct snp_guest_request_ioctl *arg,
> + enum snp_arg_type type)
You can go out to 100 characters now, so you can put "struct .. *arg" on
the top line and just put the enum on a new line.
> {
> struct snp_guest_crypto *crypto = snp_dev->crypto;
> struct snp_report_resp *resp;
> @@ -482,7 +507,7 @@ static int get_report(struct snp_guest_dev *snp_dev, struct snp_guest_request_io
> if (!arg->req_data || !arg->resp_data)
> return -EINVAL;
>
> - if (copy_from_user(&req, (void __user *)arg->req_data, sizeof(req)))
> + if (copy_from(&req, arg->req_data, sizeof(req), type))
> return -EFAULT;
>
> /*
> @@ -501,7 +526,7 @@ static int get_report(struct snp_guest_dev *snp_dev, struct snp_guest_request_io
> if (rc)
> goto e_free;
>
> - if (copy_to_user((void __user *)arg->resp_data, resp, sizeof(*resp)))
> + if (copy_to(arg->resp_data, resp, sizeof(*resp), type))
> rc = -EFAULT;
>
> e_free:
> @@ -550,7 +575,9 @@ static int get_derived_key(struct snp_guest_dev *snp_dev, struct snp_guest_reque
> return rc;
> }
>
> -static int get_ext_report(struct snp_guest_dev *snp_dev, struct snp_guest_request_ioctl *arg)
> +static int get_ext_report(struct snp_guest_dev *snp_dev,
> + struct snp_guest_request_ioctl *arg,
> + enum snp_arg_type type)
Ditto here on the 100 characters.
> {
> struct snp_guest_crypto *crypto = snp_dev->crypto;
> struct snp_ext_report_req req;
> @@ -562,7 +589,7 @@ static int get_ext_report(struct snp_guest_dev *snp_dev, struct snp_guest_reques
> if (!arg->req_data || !arg->resp_data)
> return -EINVAL;
>
> - if (copy_from_user(&req, (void __user *)arg->req_data, sizeof(req)))
> + if (copy_from(&req, arg->req_data, sizeof(req), type))
> return -EFAULT;
>
> /* userspace does not want certificate data */
> @@ -611,14 +638,13 @@ static int get_ext_report(struct snp_guest_dev *snp_dev, struct snp_guest_reques
> if (ret)
> goto e_free;
>
> - if (npages &&
> - copy_to_user((void __user *)req.certs_address, snp_dev->certs_data,
> - req.certs_len)) {
> + if (npages && copy_to(req.certs_address, snp_dev->certs_data,
> + req.certs_len, type)) {
This can also be a single line now.
Thanks,
Tom
> ret = -EFAULT;
> goto e_free;
> }
>
> - if (copy_to_user((void __user *)arg->resp_data, resp, sizeof(*resp)))
> + if (copy_to(arg->resp_data, resp, sizeof(*resp), type))
> ret = -EFAULT;
>
> e_free:
> @@ -653,13 +679,13 @@ static long snp_guest_ioctl(struct file *file, unsigned int ioctl, unsigned long
>
> switch (ioctl) {
> case SNP_GET_REPORT:
> - ret = get_report(snp_dev, &input);
> + ret = get_report(snp_dev, &input, SNP_UARG);
> break;
> case SNP_GET_DERIVED_KEY:
> ret = get_derived_key(snp_dev, &input);
> break;
> case SNP_GET_EXT_REPORT:
> - ret = get_ext_report(snp_dev, &input);
> + ret = get_ext_report(snp_dev, &input, SNP_UARG);
> break;
> default:
> break;
>
On 8/14/23 02:43, Dan Williams wrote:
> One of the common operations of a TSM (Trusted Security Module) is to
> provide a way for a TVM (confidential computing guest execution
> environment) to take a measurement of its launch state, sign it and
> submit it to a verifying party. Upon successful attestation that
> verifies the integrity of the TVM additional secrets may be deployed.
> The concept is common across TSMs, but the implementations are
> unfortunately vendor specific. While the industry grapples with a common
> definition of this attestation format [1], Linux need not make this
> problem worse by defining a new ABI per TSM that wants to perform a
> similar operation. The current momentum has been to invent new ioctl-ABI
> per TSM per function which at best is an abdication of the kernel's
> responsibility to make common infrastructure concepts share common ABI.
>
> The proposal, targeted to conceptually work with TDX, SEV, COVE if not
> more, is to define a sysfs interface to retrieve the TSM-specific blob.
>
> echo $hex_encoded_userdata_plus_nonce > /sys/class/tsm/tsm0/inhex
> hexdump /sys/class/tsm/tsm0/outblob
>
> This approach later allows for the standardization of the attestation
> blob format without needing to change the Linux ABI. Until then, the
> format of 'outblob' is determined by the parent device for 'tsm0'.
>
> The expectation is that this is a boot time exchange that need not be
> regenerated, making it amenable to a sysfs interface. In case userspace
> does try to generate multiple attestation reports it includes conflict
> detection so userspace can be sure no other thread changed the
> parameters from its last configuration step to the blob retrieval.
>
> TSM specific options are encoded as 'extra' attributes on the TSM device
> with the expectation that vendors reuse the same options for similar
> concepts. The current options are defined by SEV-SNP's need for a
> 'privilege level' concept (VMPL), and the option to retrieve a
> certificate chain in addition to the attestation report ("extended"
> format).
>
> Link: http://lore.kernel.org/r/[email protected] [1]
> Cc: Kuppuswamy Sathyanarayanan <[email protected]>
> Cc: Dionna Amalie Glaze <[email protected]>
> Cc: James Bottomley <[email protected]>
> Cc: Peter Gonda <[email protected]>
> Cc: Greg Kroah-Hartman <[email protected]>
> Cc: Samuel Ortiz <[email protected]>
> Signed-off-by: Dan Williams <[email protected]>
> ---
> Documentation/ABI/testing/sysfs-class-tsm | 47 +++++
> MAINTAINERS | 8 +
> drivers/virt/coco/Kconfig | 4
> drivers/virt/coco/Makefile | 1
> drivers/virt/coco/tdx-guest/Kconfig | 1
> drivers/virt/coco/tsm.c | 290 +++++++++++++++++++++++++++++
> include/linux/tsm.h | 45 +++++
> 7 files changed, 396 insertions(+)
> create mode 100644 Documentation/ABI/testing/sysfs-class-tsm
> create mode 100644 drivers/virt/coco/tsm.c
> create mode 100644 include/linux/tsm.h
>
> +/*
> + * Input is a small hex blob, rather than a writable binary attribute, so that
> + * it is conveyed atomically.
> + */
> +static ssize_t inhex_store(struct device *dev, struct device_attribute *attr,
> + const char *buf, size_t len)
> +{
> + u8 inblob[TSM_INBLOB_MAX];
> + size_t inblob_len;
> + int rc;
> +
> + inblob_len = len;
> + if (buf[len - 1] == '\n')
> + inblob_len--;
> + if (inblob_len & 1)
> + return -EINVAL;
> + inblob_len /= 2;
> + if (inblob_len > TSM_INBLOB_MAX)
> + return -EINVAL;
Is an array_index_nospec() call needed after this check?
> +
> + rc = hex2bin(inblob, buf, inblob_len);
> + if (rc < 0)
> + return rc;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + if (memcmp(tsm_report.desc.inblob, inblob, inblob_len) == 0)
> + return len;
> + memcpy(tsm_report.desc.inblob, inblob, inblob_len);
Should the portion of tsm_report.desc.inblob that is not updated be
cleared if inblob_len != TSM_INBLOB_MAX?
> + tsm_report.desc.inblob_len = inblob_len;
> + tsm_report.write_generation++;
> +
> + return len;
> +}
> +
> +int register_tsm(const struct tsm_ops *ops, struct device *parent,
> + const struct attribute_group **groups)
> +{
> + const struct tsm_ops *conflict;
> + struct device *dev;
> + int rc;
> +
> + if (!parent)
> + return -EINVAL;
> +
> + if (!groups)
> + groups = tsm_default_attribute_groups;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + conflict = provider.ops;
> + if (conflict) {
> + pr_err("\"%s\" ops already registered\n", conflict->name);
> + return rc;
Looks like rc is used uninitialized.
> + }
> +
> + dev = device_create_with_groups(tsm_class, parent, 0, NULL, groups,
> + "tsm0");
You can go out to 100 characters now, so this could all be one line.
Thanks,
Tom
> + if (IS_ERR(dev))
> + return PTR_ERR(dev);
> +
> + provider.ops = ops;
> + provider.dev = dev;
> + return 0;
> +}
> +EXPORT_SYMBOL_GPL(register_tsm);
> +
On Mon, Aug 14, 2023 at 11:12 AM Dan Williams <[email protected]> wrote:
>
> Jeremi Piotrowski wrote:
> > On 8/14/2023 9:43 AM, Dan Williams wrote:
> > > Changes since v1:
> > > - Switch from Keyring to sysfs (James)
> > >
> > > An attestation report is signed evidence of how a Trusted Virtual
> > > Machine (TVM) was launched and its current state. A verifying party uses
> > > the report to make judgements of the confidentiality and integrity of
> > > that execution environment. Upon successful attestation the verifying
> > > party may, for example, proceed to deploy secrets to the TVM to carry
> > > out a workload. Multiple confidential computing platforms share this
> > > similar flow.
> > >
> > > The approach of adding adding new char devs and new ioctls, for what
> > > amounts to the same logical functionality with minor formatting
> > > differences across vendors [1], is untenable. Common concepts and the
> > > community benefit from common infrastructure.
> > >
> > > Use sysfs for this facility for maintainability compared to ioctl(). The
> > > expectation is that this interface is a boot time, configure once, get
> > > report, and done flow. I.e. not something that receives ongoing
> > > transactions at runtime. However, runtime retrieval is not precluded and
> > > a mechanism to detect potential configuration conflicts from
> > > multiple-threads using this interface is included.
> > >
> >
> > I wanted to speak up to say that this does not align with the needs we have
> > in the Confidential Containers project. We want to be able to perform attestation
> > not just once during boot but during the lifecycle of the confidential VM. We
> > may need to fetch a fresh attestation report from a trusted agent but also from
> > arbitrary applications running in containers.
> >
> > The trusted agent might need attestation when launching a new container from an
> > encrypted container image or when a new secret is being added to the VM - both
> > of these events may happen at any time (also when containerized applications
> > are already executing).
> >
> > Container applications have their own uses for attestation, such as when they need
> > to fetch keys required to decrypt payloads. We also have things like performing
> > attestation when establishing a TLS or ssh connection to provide an attested e2e
> > encrypted communication channel.
>
> ...and you expect that the boot time attestation becomes invalidated
> later at run time such that ongoing round trips to the TSM are needed? I
> am looking at "Table 21. ATTESTATION_REPORT Structure" for example and
> not seeing data there that changes from one request to the next. Runtime
> validation likely looks more like the vTPM use case with PCRs. That will
> leverage the existing / common TPM ABI.
I thought we were dropping the TSM acronym as requested by Jarkko?
Why do we need to be so prescriptive about "boot time" vs "runtime"
attestations? A user may wish to attest to several requests as Jeremi
notes. And why should users be forced into using a vTPM interface if
their usecase doesn't require all the features and complexity of a
vTPM? Some users may prefer less overall code within their Trusted
Computer Base (TCB) and a TPM emulate is a significant code base.
It seems like you are just reading the SNP spec, if you read the TDX
spec you'll see there are RTMRs which can be extended with new data.
This leads to a different data in the attestation. Similar there are
REMs in the ARM CCA spec.
>
> > I don't think sysfs is suitable for such concurrent transactions. Also if you think
> > about exposing the sysfs interface to an application in a container, this requires
> > bind mounting rw part of the sysfs tree into the mount namespace - not ideal.
>
> sysfs is not suitable for concurrent transactions. The container would
> need to have an alternate path to request that the singleton owner of
> the interface generate new reports, or use the boot time attestation to
> derive per container communication sessions to the attestation agent.
On 8/14/23 02:43, Dan Williams wrote:
> One of the common operations of a TSM (Trusted Security Module) is to
> provide a way for a TVM (confidential computing guest execution
> environment) to take a measurement of its launch state, sign it and
> submit it to a verifying party. Upon successful attestation that
> verifies the integrity of the TVM additional secrets may be deployed.
> The concept is common across TSMs, but the implementations are
> unfortunately vendor specific. While the industry grapples with a common
> definition of this attestation format [1], Linux need not make this
> problem worse by defining a new ABI per TSM that wants to perform a
> similar operation. The current momentum has been to invent new ioctl-ABI
> per TSM per function which at best is an abdication of the kernel's
> responsibility to make common infrastructure concepts share common ABI.
>
> The proposal, targeted to conceptually work with TDX, SEV, COVE if not
> more, is to define a sysfs interface to retrieve the TSM-specific blob.
>
> echo $hex_encoded_userdata_plus_nonce > /sys/class/tsm/tsm0/inhex
> hexdump /sys/class/tsm/tsm0/outblob
>
> This approach later allows for the standardization of the attestation
> blob format without needing to change the Linux ABI. Until then, the
> format of 'outblob' is determined by the parent device for 'tsm0'.
>
> The expectation is that this is a boot time exchange that need not be
> regenerated, making it amenable to a sysfs interface. In case userspace
> does try to generate multiple attestation reports it includes conflict
> detection so userspace can be sure no other thread changed the
> parameters from its last configuration step to the blob retrieval.
>
> TSM specific options are encoded as 'extra' attributes on the TSM device
> with the expectation that vendors reuse the same options for similar
> concepts. The current options are defined by SEV-SNP's need for a
> 'privilege level' concept (VMPL), and the option to retrieve a
> certificate chain in addition to the attestation report ("extended"
> format).
>
> Link: http://lore.kernel.org/r/[email protected] [1]
> Cc: Kuppuswamy Sathyanarayanan <[email protected]>
> Cc: Dionna Amalie Glaze <[email protected]>
> Cc: James Bottomley <[email protected]>
> Cc: Peter Gonda <[email protected]>
> Cc: Greg Kroah-Hartman <[email protected]>
> Cc: Samuel Ortiz <[email protected]>
> Signed-off-by: Dan Williams <[email protected]>
> ---
> Documentation/ABI/testing/sysfs-class-tsm | 47 +++++
> MAINTAINERS | 8 +
> drivers/virt/coco/Kconfig | 4
> drivers/virt/coco/Makefile | 1
> drivers/virt/coco/tdx-guest/Kconfig | 1
> drivers/virt/coco/tsm.c | 290 +++++++++++++++++++++++++++++
> include/linux/tsm.h | 45 +++++
> 7 files changed, 396 insertions(+)
> create mode 100644 Documentation/ABI/testing/sysfs-class-tsm
> create mode 100644 drivers/virt/coco/tsm.c
> create mode 100644 include/linux/tsm.h
>
> +static ssize_t privlevel_store(struct device *dev,
> + struct device_attribute *attr, const char *buf,
> + size_t len)
> +{
> + unsigned int val;
> + int rc;
> +
> + rc = kstrtouint(buf, 0, &val);
> + if (rc)
> + return rc;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + if (tsm_report.desc.privlevel == val)
> + return len;
> + tsm_report.desc.privlevel = val;
> + tsm_report.write_generation++;
So I'm wondering if this use of write_generation helps or not. Since it
isn't incremented if the levels are the same, that might present race
conditions.
What if user A requests vmpl 2 and privlevel is already 2, then
write_generation is not incremented. But before user A can read back the
generation value user B can request vmpl 3 and cause write_generation to
be incremented.
This may not be a problem for VMPL, since that can be checked in the
returned attestation report, but it could be for the report format. If the
extended format is requested but changed to default, then the additional
certs might not be returned and the guest may think there aren't any...?
Maybe incrementing the write_generation no matter what is best.
Thanks,
Tom
> +
> + return len;
> +}
> +
> +static ssize_t privlevel_show(struct device *dev, struct device_attribute *attr,
> + char *buf)
> +{
> + return sysfs_emit(buf, "%u\n", tsm_report.desc.privlevel);
> +}
> +
> +static DEVICE_ATTR_RW(privlevel);
> +
> +static ssize_t format_store(struct device *dev, struct device_attribute *attr,
> + const char *buf, size_t len)
> +{
> + enum tsm_format format;
> +
> + if (sysfs_streq(buf, "default"))
> + format = TSM_FORMAT_DEFAULT;
> + else if (sysfs_streq(buf, "extended"))
> + format = TSM_FORMAT_EXTENDED;
> + else
> + return -EINVAL;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + if (tsm_report.desc.outblob_format == format)
> + return len;
> + tsm_report.desc.outblob_format = format;
> + tsm_report.write_generation++;
> +
> + return len;
> +}
> +
> +static ssize_t format_show(struct device *dev, struct device_attribute *attr,
> + char *buf)
> +{
> + if (tsm_report.desc.outblob_format == TSM_FORMAT_DEFAULT)
> + return sysfs_emit(buf, "default\n");
> + return sysfs_emit(buf, "extended\n");
> +}
> +
> +static DEVICE_ATTR_RW(format);
> +
> +static struct attribute *tsm_extra_attributes[] = {
> + &dev_attr_format.attr,
> + &dev_attr_privlevel.attr,
> + NULL,
> +};
> +
> +struct attribute_group tsm_extra_attribute_group = {
> + .attrs = tsm_extra_attributes,
> +};
> +EXPORT_SYMBOL_GPL(tsm_extra_attribute_group);
> +
> +/*
> + * Input is a small hex blob, rather than a writable binary attribute, so that
> + * it is conveyed atomically.
> + */
> +static ssize_t inhex_store(struct device *dev, struct device_attribute *attr,
> + const char *buf, size_t len)
> +{
> + u8 inblob[TSM_INBLOB_MAX];
> + size_t inblob_len;
> + int rc;
> +
> + inblob_len = len;
> + if (buf[len - 1] == '\n')
> + inblob_len--;
> + if (inblob_len & 1)
> + return -EINVAL;
> + inblob_len /= 2;
> + if (inblob_len > TSM_INBLOB_MAX)
> + return -EINVAL;
> +
> + rc = hex2bin(inblob, buf, inblob_len);
> + if (rc < 0)
> + return rc;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + if (memcmp(tsm_report.desc.inblob, inblob, inblob_len) == 0)
> + return len;
> + memcpy(tsm_report.desc.inblob, inblob, inblob_len);
> + tsm_report.desc.inblob_len = inblob_len;
> + tsm_report.write_generation++;
> +
> + return len;
> +}
> +
> +static ssize_t inhex_show(struct device *dev, struct device_attribute *attr,
> + char *buf)
> +{
> + char *end;
> +
> + guard(rwsem_read)(&tsm_rwsem);
> + if (!tsm_report.desc.inblob_len)
> + return 0;
> + end = bin2hex(buf, tsm_report.desc.inblob, tsm_report.desc.inblob_len);
> + *end++ = '\n';
> + return end - buf;
> +}
> +static DEVICE_ATTR_RW(inhex);
> +
> +static ssize_t generation_show(struct device *dev,
> + struct device_attribute *attr, char *buf)
> +{
> + guard(rwsem_read)(&tsm_rwsem);
> + return sysfs_emit(buf, "%lu\n", tsm_report.write_generation);
> +}
> +static DEVICE_ATTR_RO(generation);
> +
> +static struct attribute *tsm_attributes[] = {
> + &dev_attr_inhex.attr,
> + &dev_attr_generation.attr,
> + NULL,
> +};
> +
> +static ssize_t outblob_read(struct file *f, struct kobject *kobj,
> + struct bin_attribute *bin_attr, char *buf,
> + loff_t offset, size_t count)
> +{
> + guard(rwsem_read)(&tsm_rwsem);
> + if (!tsm_report.desc.inblob_len)
> + return -EINVAL;
> +
> + if (!tsm_report.outblob ||
> + tsm_report.read_generation != tsm_report.write_generation) {
> + const struct tsm_ops *ops = provider.ops;
> + size_t outblob_len;
> + u8 *outblob;
> +
> + kvfree(tsm_report.outblob);
> + outblob = ops->report_new(provider.dev->parent,
> + &tsm_report.desc, &outblob_len);
> + if (IS_ERR(outblob))
> + return PTR_ERR(outblob);
> + tsm_report.outblob_len = outblob_len;
> + tsm_report.outblob = outblob;
> + tsm_report.read_generation = tsm_report.write_generation;
> + }
> +
> + return memory_read_from_buffer(buf, count, &offset,
> + tsm_report.outblob,
> + tsm_report.outblob_len);
> +}
> +static BIN_ATTR_RO(outblob, 0);
> +
> +static struct bin_attribute *tsm_bin_attributes[] = {
> + &bin_attr_outblob,
> + NULL,
> +};
> +
> +struct attribute_group tsm_default_attribute_group = {
> + .bin_attrs = tsm_bin_attributes,
> + .attrs = tsm_attributes,
> +};
> +EXPORT_SYMBOL_GPL(tsm_default_attribute_group);
> +
> +static const struct attribute_group *tsm_default_attribute_groups[] = {
> + &tsm_default_attribute_group,
> + NULL,
> +};
> +
> +int register_tsm(const struct tsm_ops *ops, struct device *parent,
> + const struct attribute_group **groups)
> +{
> + const struct tsm_ops *conflict;
> + struct device *dev;
> + int rc;
> +
> + if (!parent)
> + return -EINVAL;
> +
> + if (!groups)
> + groups = tsm_default_attribute_groups;
> +
> + guard(rwsem_write)(&tsm_rwsem);
> + conflict = provider.ops;
> + if (conflict) {
> + pr_err("\"%s\" ops already registered\n", conflict->name);
> + return rc;
> + }
> +
> + dev = device_create_with_groups(tsm_class, parent, 0, NULL, groups,
> + "tsm0");
> + if (IS_ERR(dev))
> + return PTR_ERR(dev);
> +
> + provider.ops = ops;
> + provider.dev = dev;
> + return 0;
> +}
> +EXPORT_SYMBOL_GPL(register_tsm);
> +
> +int unregister_tsm(const struct tsm_ops *ops)
> +{
> + guard(rwsem_write)(&tsm_rwsem);
> + if (ops != provider.ops)
> + return -EBUSY;
> + provider.ops = NULL;
> + device_unregister(provider.dev);
> + provider.dev = NULL;
> + kvfree(tsm_report.outblob);
> + tsm_report.outblob = NULL;
> + return 0;
> +}
> +EXPORT_SYMBOL_GPL(unregister_tsm);
> +
> +static int __init tsm_init(void)
> +{
> + tsm_class = class_create("tsm");
> + return PTR_ERR_OR_ZERO(tsm_class);
> +}
> +module_init(tsm_init);
> +
> +static void __exit tsm_exit(void)
> +{
> + class_destroy(tsm_class);
> +}
> +module_exit(tsm_exit);
> +
> +MODULE_LICENSE("GPL");
> +MODULE_DESCRIPTION("Provide Trusted Security Module attestation reports via sysfs");
> diff --git a/include/linux/tsm.h b/include/linux/tsm.h
> new file mode 100644
> index 000000000000..6dc2f07543b8
> --- /dev/null
> +++ b/include/linux/tsm.h
> @@ -0,0 +1,45 @@
> +/* SPDX-License-Identifier: GPL-2.0 */
> +#ifndef __TSM_H
> +#define __TSM_H
> +
> +#include <linux/types.h>
> +#include <linux/device.h>
> +
> +#define TSM_INBLOB_MAX 64
> +
> +enum tsm_format {
> + TSM_FORMAT_DEFAULT,
> + TSM_FORMAT_EXTENDED,
> +};
> +
> +/**
> + * struct tsm_desc - option descriptor for generating tsm report blobs
> + * @privlevel: optional privilege level to associate with @outblob
> + * @inblob_len: sizeof @inblob
> + * @inblob: arbitrary input data
> + * @outblob_format: for TSMs with an "extended" format
> + */
> +struct tsm_desc {
> + unsigned int privlevel;
> + size_t inblob_len;
> + u8 inblob[TSM_INBLOB_MAX];
> + enum tsm_format outblob_format;
> +};
> +
> +/*
> + * arch specific ops, only one is expected to be registered at a time
> + * i.e. only one of SEV, TDX, COVE, etc.
> + */
> +struct tsm_ops {
> + const char *name;
> + u8 *(*report_new)(struct device *dev, const struct tsm_desc *desc,
> + size_t *outblob_len);
> +};
> +
> +extern struct attribute_group tsm_default_attribute_group;
> +extern struct attribute_group tsm_extra_attribute_group;
> +
> +int register_tsm(const struct tsm_ops *ops, struct device *parent,
> + const struct attribute_group **groups);
> +int unregister_tsm(const struct tsm_ops *ops);
> +#endif /* __TSM_H */
>
>
On Mon, Aug 14, 2023 at 11:04:52AM +0200, Jeremi Piotrowski wrote:
Good morning, I hope the week is going well for everyone.
> On 8/14/2023 9:43 AM, Dan Williams wrote:
> > Changes since v1:
> > - Switch from Keyring to sysfs (James)
> >
> > An attestation report is signed evidence of how a Trusted Virtual
> > Machine (TVM) was launched and its current state. A verifying party uses
> > the report to make judgements of the confidentiality and integrity of
> > that execution environment. Upon successful attestation the verifying
> > party may, for example, proceed to deploy secrets to the TVM to carry
> > out a workload. Multiple confidential computing platforms share this
> > similar flow.
> >
> > The approach of adding adding new char devs and new ioctls, for what
> > amounts to the same logical functionality with minor formatting
> > differences across vendors [1], is untenable. Common concepts and the
> > community benefit from common infrastructure.
> >
> > Use sysfs for this facility for maintainability compared to ioctl(). The
> > expectation is that this interface is a boot time, configure once, get
> > report, and done flow. I.e. not something that receives ongoing
> > transactions at runtime. However, runtime retrieval is not precluded and
> > a mechanism to detect potential configuration conflicts from
> > multiple-threads using this interface is included.
> >
> I wanted to speak up to say that this does not align with the needs
> we have in the Confidential Containers project. We want to be able
> to perform attestation not just once during boot but during the
> lifecycle of the confidential VM. We may need to fetch a fresh
> attestation report from a trusted agent but also from arbitrary
> applications running in containers.
>
> The trusted agent might need attestation when launching a new
> container from an encrypted container image or when a new secret is
> being added to the VM - both of these events may happen at any time
> (also when containerized applications are already executing).
>
> Container applications have their own uses for attestation, such as
> when they need to fetch keys required to decrypt payloads. We also
> have things like performing attestation when establishing a TLS or
> ssh connection to provide an attested e2e encrypted communication
> channel.
>
> I don't think sysfs is suitable for such concurrent
> transactions. Also if you think about exposing the sysfs interface
> to an application in a container, this requires bind mounting rw
> part of the sysfs tree into the mount namespace - not ideal.
We don't have a dog in this fight regarding TDX [1], but we do have a
significant body of experience with the concepts and challenges
involved.
The issue at hand is that trust is a resource that needs to be
orchestrated, just like any other resource. A concept, based on our
experiences, that seems to be significantly outside of mainstream
thought.
The notion of the need to orchestrate trust seems to be particularly
important with a concept such as Confidential Containers.
FWIW, we have pushed forward a second round of patches for the kernel
infrastructure that make the concepts of trust orchestration and
containerization tenable:
https://lore.kernel.org/linux-security-module/[email protected]/T/#t
In addition, FWIW, we have actually built systems that implement these
principals, obviously not on TDX hardware, see [1], but we do have a
significant body of experience with using SGX as a trust root. We
have even provided a substantial set of initial userspace tooling that
implement these concepts to support the proposed patches.
I only say all of this to convey the notion that we have actually done
work on all of these concepts and are not just waving our hands
around.
The notion of surfacing this information through /sysfs becomes less
problematic if one approaches the issue through the lens of having
trust orchestrators that are responsible for managing the security or
trust status of the execution platform at large and any subordinate
workloads.
This concept is true, even if the platform/VM is only hosting a single
workload. Which may be a necessity for some security contracts, where
there is literally no trust in the fact that side-channel disclosure
threats can be properly mitigated, ie. there cannot be workloads with
possible adversarial intents.
From the outside looking in, unless there are some fundamental
conversations regarding how trusted systems and workloads can be
developed with the architectures being proposed, it is completely
unclear how durable API's, on the order of 20+ years are ever going to
be attained.
No criticism, just observation.
> Jeremi
Best wishes for continued progress on all of this, it is important
stuff, both in the cloud and on the edge.
As always,
Dr. Greg
The Quixote Project - Flailing at the Travails of Cybersecurity
[1]: TDX enabled hardware is difficult, if not impossible to obtain.
If there is any doubt, simply search for TDX hardware availability and
cringe at the conversations on the Intel forums about people trying to
get experience with the technology.
This is the same problem that plagued TXT and SGX and results in
enabling infrastructure development being done in an echo chamber. A
concept that may have proven successful when all of this work was
expected to be implementated and enabled by OEM providers but seems
problematic in an ostensibly 'community' driven project such as Linux.
On Wed, Aug 23, 2023 at 03:49:17PM +0200, Samuel Ortiz wrote:
Good morning, I hope the week is starting well for everyone.
Some further background to hopefully assist decision making on how to
handle the export of attestation information for COCO TSM's.
> Hi Dan,
>
> On Mon, Aug 14, 2023 at 12:43:21AM -0700, Dan Williams wrote:
> > One of the common operations of a TSM (Trusted Security Module) is to
> > provide a way for a TVM (confidential computing guest execution
> > environment) to take a measurement of its launch state, sign it and
> > submit it to a verifying party. Upon successful attestation that
> > verifies the integrity of the TVM additional secrets may be deployed.
> > The concept is common across TSMs, but the implementations are
> > unfortunately vendor specific. While the industry grapples with a common
> > definition of this attestation format [1], Linux need not make this
> > problem worse by defining a new ABI per TSM that wants to perform a
> > similar operation. The current momentum has been to invent new ioctl-ABI
> > per TSM per function which at best is an abdication of the kernel's
> > responsibility to make common infrastructure concepts share common ABI.
> >
> > The proposal, targeted to conceptually work with TDX, SEV, COVE if not
> > more, is to define a sysfs interface to retrieve the TSM-specific blob.
> >
> > echo $hex_encoded_userdata_plus_nonce > /sys/class/tsm/tsm0/inhex
> > hexdump /sys/class/tsm/tsm0/outblob
> My concern with that interface is that one could easily get an
> attestation report with a nonce set by another userspace component
> or thread. I realize there is a generation counter to detect if a
> configuration changed between the caller's last config setting and
> the report it got, but I think that this shows that this may not be
> the best interface. IMHO an attestation report request from
> userspace should be an atomic call that includes multiple platform
> independent attibutes like e.g. an attestation nonce.
The challenge in all of this would seem to be the need to get a 20+
year ABI 'right' on the first try, so some issues to consider based on
our experiences building trusted systems, albeit with a focus on
endpoint devices.
The issue about needing to implement attestation atomically is on
point. This requirement would seem to suggest that the optimum path
forward for confidential computing will be to have attestation support
implemented in the resource orchestration infrastructure inside the
TVM.
Since the goal is confidentiality, you need to trust everything that
executes inside the execution domain/VM. This is particularly true
given the reality of micro-architectural side channel issues, which
don't seem to be going away and which Intel explicitly documents that
you must address with TDX.
Given that, there would seem to be no reason not to have the
attestation support all in one place and then provide access to that
functionality to anything in the TVM that would want to make
an integrity statement to a relying party.
A sysfs interface is perfectly suited to the this model, since the
trust orchestrator, if we use our terminology, can address the
concurrency and atomicity issues presented by a sysfs interface.
From an Intel/SGX centric perspective, which is our area of expertise
when it comes to the notion of a TSM, the attestation process is
reasonably complex, even more so in a TDX environment, since you need
to go outside the trusted execution domain in order to conduct the
full attestation process. I don't know how many people have direct
experience with all of this, but we can offer it on good authority
that the COCO model will benefit from having all the necessary
procedures done correctly and in one place
> > This approach later allows for the standardization of the
> > attestation blob format without needing to change the Linux
> > ABI. Until then, the format of 'outblob' is determined by the
> > parent device for 'tsm0'.
> >
> > The expectation is that this is a boot time exchange that need not
> > be regenerated,
> This works well with the encrypted boot disk that's decrypted
> through attestation use-case, but this is just one use
> case. Although I don't expect attestation requests to be frequent,
> we should not assume this is only a boot time operation. Not only it
> can happen after the guest is fully booted, but it can also happen
> multiple times. An attestation flow where a guest gets an
> attestation token back from a validated report is something we'd
> want to support. Those token's validity are time limited, and
> userspace would want to regenerate a report, with a fresh,
> attestation service provided nonce.
There would certainly seem to be no argument that a trusted execution
domain may want to conduct multiple attestations, Jeremi suggested
this is something the Confidential Containers initiative would want as
well.
The model of having attestation support centralized in the resource
orchestration infrastructure is not only consistent with that need but
would also appear to be the optimal approach.
Let us consider TDX as an example, since it will undoubtedly be a
major player in COCO.
In the Data Center Attestation Primitives (DCAP) model, one needs to
choregraph a dance between the Provisioning Certification Enclave
(PCE) and the Quoting Enclave (QE) in order to conduct the attestation
process. That requires access to SGX functionality, something that is
not supported by Intel in a TDX mediated virtual machine.
This imposes a need to go outside of the TD guest, to the host, in
order to get access to SGX in order to load and run the PCE and QE
enclaves.
The call into the kernel to access the hypervisor interface is simply
the starting point for the attestation process. The purpose is to
have the TDX module generate a CPU specific report (SEAMREPORT) that
will ultimately authenticate material, usually ECDH key exchange
components, as coming from an execution domain with known provenance.
That report then needs to be processed through the PCE and QE enclave
infrastructure in order to fully prove the identity of the report and
thus authenticate its contents. Mixed into all of this is a need to
verify the MROWNER and MRSIGNER signatures on the involved enclaves as
well as the Security Version Numbers (SVN's) on the enclaves and the
TDX infrastructure. This process ends up authenticating a certificate
chain which proves that a secret to be conveyed into the TD is
actually being injected into a valid TD, with known provenance and one
that is 'Genuine Intel' in origin.
All of this doesn't seem to speak of infrastructure you would want to
replicate multiple times over.
> Another thing to keep in mind is that an attestation report could be
> amended by userspace itself, for TEE that support runtime
> measurement (The RTMR things...). So the TVM measurement itself
> could change during the lifecycle of a TVM.
Once again, speaking only to TDX, it isn't apparent that the RTMR
registers are going to be useful for the task of confirming general
application level integrity once the OS is started. Their primary
utility will be to provide a root of trust for the resource
orchestration infrastructure and in turn VTPM's, or something like
TSEM's root security modeling namespace, if the goal is to provide a
foundation for generalized trust/integrity mechanisms inside of the
TVM.
To wit:
The formal usage recommendations for the TDX RTMR registers are as follows:
RTMR[0]: PCR[1,7]
RTMR[1]: PCR[4,5]
RTMR[2]: PCR[8,15]
RTMR[3]: Special use
Where RTMR[2] is specified for usage by the OS/application for
measurement purposes, so only the equivalent of one PCR is available.
This would automatically preclude the use of systemd in a 'trusted' VM
implementation, given that the latest Linux TPM PCR registry has
systemd consuming PCR registers 11-15 for its own use. Once again, if
you are really serious about confidentiality, you need to have an
attestable integrity guarantee for everything that happens inside the
trusted VM. IMA uses register 10 if you go that route and we will be
using a register above what systemd is using for the TSEM root
namespace.
So we believe the lone trusted virtual machine measurement register
will need to stay constant, in order to easily validate whatever root
of trust that is in turn designated to guarantee the integrity of what
goes on inside of the TVM from an OS/application perspective.
A functional attestation scheme is thus going to require not only
attestation of the state of the TVM but also what has gone on inside
of the VM after it has booted. Another reason for having a single
shopping site for the attestation needs of whatever runs inside of the
trusted VM.
FWIW, since we copied the LSM list on this, we haven't come by the
design and implementation of TSEM and the notion of trust
orchestration idly. If COCO is going to become a relevant reality
there needs to be consideration given to the concept of trust
orchestraton and management.
Once gain, FWIW, and something that Jeremi from Microsoft might be
able to comment on. It is unclear how COCO based Confidential
Containers fits into the standard IMA attestation model, where one has
to review an event log documented by a linear extension trust
measurement to determine whether or not the system is to be trusted.
At a minimum, this would require that a relying party for attestation
of one container be able to verify what has gone on in all other
container invocations. Unless the containers are being run for only a
single entity, this would seem to imply a violation of the notion of
privacy and confidentiality.
> Cheers,
> Samuel.
Hopefully all of the above is of assistance in getting the ABI right
and the scope of issues involved.
Best wishes for a productive week.
As always,
Dr. Greg
The Quixote Project - Flailing at the Travails of Cybersecurity