Skip to contentSkip to conversation

kvm: s390: Add PV dump support

PATCH Superseded Janosch Frank 3 participants 26 messages May 16, 2022 – May 17, 2022

Subsystems (4)

v5 of 6 Superseded Latest is v6 →

lore
Share & tools
lore

Janosch Frank

author maintainer
#0 lore raw

v5 cover kvm: s390: Add PV dump support

Sometimes dumping inside of a VM fails, is unavailable or doesn't yield the required data. For these occasions we dump the VM from the outside, writing memory and cpu data to a file.

Up to now PV guests only supported dumping from the inside of the guest through dumpers like KDUMP. A PV guest can be dumped from the hypervisor but the data will be stale and / or encrypted.

To get the actual state of the PV VM we need the help of the Ultravisor who safeguards the VM state. New UV calls have been added to initialize the dump, dump storage state data, dump cpu data and complete the dump process.

I chose not to document the dump data provided by the Ultravisor since KVM doesn't interprete it in any way. We're currently searching for a location and enough cycles to make it available to all.

v5:
* v4 went out without a version
* Fixed spelling mistake in query
* Improved the storage state dump code
* Switched to the scheduling uv call for all dump commands
* Fixed the documentation formatting
* Moved the capability into its own patch it shouldn't be part of patch #6

v4:
* Rebased and fixed up conflicts due to the Documentation changes and new KVM capabilities
* Fixed the dump facility check, now we check for all 4 calls

v3:
* Added Rev-by
* Renamed the query function's len variables to len_min

v2:
* Added vcpu SIE blocking to avoid validities
* Moved the KVM CAP to patch #7
* Renamed len to len_max and introduced len_written for extendability
* Added Rev-bys

Janosch Frank (10):
s390x: Add SE hdr query information
s390: uv: Add dump fields to query
KVM: s390: pv: Add query interface
KVM: s390: pv: Add dump support definitions
KVM: s390: pv: Add query dump information
kvm: s390: Add configuration dump functionality
kvm: s390: Add CPU dump functionality
kvm: s390: Add KVM_CAP_S390_PROTECTED_DUMP
Documentation: virt: Protected virtual machine dumps Documentation/virt/kvm/api.rst: Add protvirt dump/info api
descriptions

 Documentation/virt/kvm/api.rst               | 153 +++++++++-
 Documentation/virt/kvm/s390/index.rst        |   1 +
 Documentation/virt/kvm/s390/s390-pv-dump.rst |  64 ++++
 arch/s390/boot/uv.c                          |   4 +
 arch/s390/include/asm/kvm_host.h             |   1 +
 arch/s390/include/asm/uv.h                   |  45 ++-
 arch/s390/kernel/uv.c                        |  53 ++++
 arch/s390/kvm/kvm-s390.c                     | 306 +++++++++++++++++++
 arch/s390/kvm/kvm-s390.h                     |   3 +
 arch/s390/kvm/pv.c                           | 129 ++++++++
 include/uapi/linux/kvm.h                     |  55 ++++
 11 files changed, 811 insertions(+), 3 deletions(-)
 create mode 100644 Documentation/virt/kvm/s390/s390-pv-dump.rst
Signature · 1 line
2.34.1
author maintainer
#1 lore raw

PATCH v5 1/10 s390x: Add SE hdr query information

We have information about the supported se header version and pcf bits so let's expose it via the sysfs files.

---
 arch/s390/boot/uv.c        |  2 ++
 arch/s390/include/asm/uv.h |  7 ++++++-
 arch/s390/kernel/uv.c      | 20 ++++++++++++++++++++
 3 files changed, 28 insertions(+), 1 deletion(-)
arch/s390/boot/uv.c +2 -0
index e6be155ab2e5..b100b57cf15d 100644
@@ -41,6 +41,8 @@ void uv_query_info(void)
 		uv_info.max_num_sec_conf = uvcb.max_num_sec_conf;
 		uv_info.max_guest_cpu_id = uvcb.max_guest_cpu_id;
 		uv_info.uv_feature_indications = uvcb.uv_feature_indications;
+		uv_info.supp_se_hdr_ver = uvcb.supp_se_hdr_versions;
+		uv_info.supp_se_hdr_pcf = uvcb.supp_se_hdr_pcf;
 	}
 
 #ifdef CONFIG_PROTECTED_VIRTUALIZATION_GUEST
arch/s390/include/asm/uv.h +6 -1
index cfea7b77a5b8..46498b8c587b 100644
@@ -110,7 +110,10 @@ struct uv_cb_qui {
 	u8  reserved88[158 - 136];		/* 0x0088 */
 	u16 max_guest_cpu_id;			/* 0x009e */
 	u64 uv_feature_indications;		/* 0x00a0 */
-	u8  reserveda8[200 - 168];		/* 0x00a8 */
+	u64 reserveda8;				/* 0x00a8 */
+	u64 supp_se_hdr_versions;		/* 0x00b0 */
+	u64 supp_se_hdr_pcf;			/* 0x00b8 */
+	u64 reservedc0;				/* 0x00c0 */
 } __packed __aligned(8);
 
 /* Initialize Ultravisor */
@@ -307,6 +310,8 @@ struct uv_info {
 	unsigned int max_num_sec_conf;
 	unsigned short max_guest_cpu_id;
 	unsigned long uv_feature_indications;
+	unsigned long supp_se_hdr_ver;
+	unsigned long supp_se_hdr_pcf;
 };
 
 extern struct uv_info uv_info;
arch/s390/kernel/uv.c +20 -0
index a5425075dd25..852840384e75 100644
@@ -392,6 +392,24 @@ static ssize_t uv_query_facilities(struct kobject *kobj,
 static struct kobj_attribute uv_query_facilities_attr =
 	__ATTR(facilities, 0444, uv_query_facilities, NULL);
 
+static ssize_t uv_query_supp_se_hdr_ver(struct kobject *kobj,
+					struct kobj_attribute *attr, char *buf)
+{
+	return sysfs_emit(buf, "%lx\n", uv_info.supp_se_hdr_ver);
+}
+
+static struct kobj_attribute uv_query_supp_se_hdr_ver_attr =
+	__ATTR(supp_se_hdr_ver, 0444, uv_query_supp_se_hdr_ver, NULL);
+
+static ssize_t uv_query_supp_se_hdr_pcf(struct kobject *kobj,
+					struct kobj_attribute *attr, char *buf)
+{
+	return sysfs_emit(buf, "%lx\n", uv_info.supp_se_hdr_pcf);
+}
+
+static struct kobj_attribute uv_query_supp_se_hdr_pcf_attr =
+	__ATTR(supp_se_hdr_pcf, 0444, uv_query_supp_se_hdr_pcf, NULL);
+
 static ssize_t uv_query_feature_indications(struct kobject *kobj,
 					    struct kobj_attribute *attr, char *buf)
 {
@@ -437,6 +455,8 @@ static struct attribute *uv_query_attrs[] = {
 	&uv_query_max_guest_cpus_attr.attr,
 	&uv_query_max_guest_vms_attr.attr,
 	&uv_query_max_guest_addr_attr.attr,
+	&uv_query_supp_se_hdr_ver_attr.attr,
+	&uv_query_supp_se_hdr_pcf_attr.attr,
 	NULL,
 };
 
Signature · 1 line
2.34.1
author maintainer
#3 lore raw

PATCH v5 2/10 s390: uv: Add dump fields to query

In reply to Janosch Frank

The new dump feature requires us to know how much memory is needed for the "dump storage state" and "dump finalize" ultravisor call. These values are reported via the UV query call.

---
 arch/s390/boot/uv.c        |  2 ++
 arch/s390/include/asm/uv.h |  5 +++++
 arch/s390/kernel/uv.c      | 33 +++++++++++++++++++++++++++++++++
 3 files changed, 40 insertions(+)
arch/s390/boot/uv.c +2 -0
index b100b57cf15d..67c737c1e580 100644
@@ -43,6 +43,8 @@ void uv_query_info(void)
 		uv_info.uv_feature_indications = uvcb.uv_feature_indications;
 		uv_info.supp_se_hdr_ver = uvcb.supp_se_hdr_versions;
 		uv_info.supp_se_hdr_pcf = uvcb.supp_se_hdr_pcf;
+		uv_info.conf_dump_storage_state_len = uvcb.conf_dump_storage_state_len;
+		uv_info.conf_dump_finalize_len = uvcb.conf_dump_finalize_len;
 	}
 
 #ifdef CONFIG_PROTECTED_VIRTUALIZATION_GUEST
arch/s390/include/asm/uv.h +5 -0
index 46498b8c587b..e8257a293dd1 100644
@@ -114,6 +114,9 @@ struct uv_cb_qui {
 	u64 supp_se_hdr_versions;		/* 0x00b0 */
 	u64 supp_se_hdr_pcf;			/* 0x00b8 */
 	u64 reservedc0;				/* 0x00c0 */
+	u64 conf_dump_storage_state_len;	/* 0x00c8 */
+	u64 conf_dump_finalize_len;		/* 0x00d0 */
+	u8  reservedd8[256 - 216];		/* 0x00d8 */
 } __packed __aligned(8);
 
 /* Initialize Ultravisor */
@@ -312,6 +315,8 @@ struct uv_info {
 	unsigned long uv_feature_indications;
 	unsigned long supp_se_hdr_ver;
 	unsigned long supp_se_hdr_pcf;
+	unsigned long conf_dump_storage_state_len;
+	unsigned long conf_dump_finalize_len;
 };
 
 extern struct uv_info uv_info;
arch/s390/kernel/uv.c +33 -0
index 852840384e75..84fe33b6af4d 100644
@@ -410,6 +410,36 @@ static ssize_t uv_query_supp_se_hdr_pcf(struct kobject *kobj,
 static struct kobj_attribute uv_query_supp_se_hdr_pcf_attr =
 	__ATTR(supp_se_hdr_pcf, 0444, uv_query_supp_se_hdr_pcf, NULL);
 
+static ssize_t uv_query_dump_cpu_len(struct kobject *kobj,
+				     struct kobj_attribute *attr, char *page)
+{
+	return scnprintf(page, PAGE_SIZE, "%lx\n",
+			uv_info.guest_cpu_stor_len);
+}
+
+static struct kobj_attribute uv_query_dump_cpu_len_attr =
+	__ATTR(uv_query_dump_cpu_len, 0444, uv_query_dump_cpu_len, NULL);
+
+static ssize_t uv_query_dump_storage_state_len(struct kobject *kobj,
+					       struct kobj_attribute *attr, char *page)
+{
+	return scnprintf(page, PAGE_SIZE, "%lx\n",
+			uv_info.conf_dump_storage_state_len);
+}
+
+static struct kobj_attribute uv_query_dump_storage_state_len_attr =
+	__ATTR(dump_storage_state_len, 0444, uv_query_dump_storage_state_len, NULL);
+
+static ssize_t uv_query_dump_finalize_len(struct kobject *kobj,
+					  struct kobj_attribute *attr, char *page)
+{
+	return scnprintf(page, PAGE_SIZE, "%lx\n",
+			uv_info.conf_dump_finalize_len);
+}
+
+static struct kobj_attribute uv_query_dump_finalize_len_attr =
+	__ATTR(dump_finalize_len, 0444, uv_query_dump_finalize_len, NULL);
+
 static ssize_t uv_query_feature_indications(struct kobject *kobj,
 					    struct kobj_attribute *attr, char *buf)
 {
@@ -457,6 +487,9 @@ static struct attribute *uv_query_attrs[] = {
 	&uv_query_max_guest_addr_attr.attr,
 	&uv_query_supp_se_hdr_ver_attr.attr,
 	&uv_query_supp_se_hdr_pcf_attr.attr,
+	&uv_query_dump_storage_state_len_attr.attr,
+	&uv_query_dump_finalize_len_attr.attr,
+	&uv_query_dump_cpu_len_attr.attr,
 	NULL,
 };
 
Signature · 1 line
2.34.1
author maintainer
#5 lore raw

PATCH v5 3/10 KVM: s390: pv: Add query interface

In reply to Janosch Frank

Some of the query information is already available via sysfs but having a IOCTL makes the information easier to retrieve.

---
 arch/s390/kvm/kvm-s390.c | 76 ++++++++++++++++++++++++++++++++++++++++
 include/uapi/linux/kvm.h | 25 +++++++++++++
 2 files changed, 101 insertions(+)
arch/s390/kvm/kvm-s390.c +76 -0
index 76ad6408cb2c..5859f243d287 100644
@@ -2224,6 +2224,42 @@ static int kvm_s390_cpus_to_pv(struct kvm *kvm, u16 *rc, u16 *rrc)
 	return r;
 }
 
+/*
+ * Here we provide user space with a direct interface to query UV
+ * related data like UV maxima and available features as well as
+ * feature specific data.
+ *
+ * To facilitate future extension of the data structures we'll try to
+ * write data up to the maximum requested length.
+ */
+static ssize_t kvm_s390_handle_pv_info(struct kvm_s390_pv_info *info)
+{
+	ssize_t len_min;
+
+	switch (info->header.id) {
+	case KVM_PV_INFO_VM: {
+		len_min =  sizeof(info->header) + sizeof(info->vm);
+
+		if (info->header.len_max < len_min)
+			return -EINVAL;
+
+		memcpy(info->vm.inst_calls_list,
+		       uv_info.inst_calls_list,
+		       sizeof(uv_info.inst_calls_list));
+
+		/* It's max cpuid not max cpus, so it's off by one */
+		info->vm.max_cpus = uv_info.max_guest_cpu_id + 1;
+		info->vm.max_guests = uv_info.max_num_sec_conf;
+		info->vm.max_guest_addr = uv_info.max_sec_stor_addr;
+		info->vm.feature_indication = uv_info.uv_feature_indications;
+
+		return len_min;
+	}
+	default:
+		return -EINVAL;
+	}
+}
+
 static int kvm_s390_handle_pv(struct kvm *kvm, struct kvm_pv_cmd *cmd)
 {
 	int r = 0;
@@ -2360,6 +2396,46 @@ static int kvm_s390_handle_pv(struct kvm *kvm, struct kvm_pv_cmd *cmd)
 			     cmd->rc, cmd->rrc);
 		break;
 	}
+	case KVM_PV_INFO: {
+		struct kvm_s390_pv_info info = {};
+		ssize_t data_len;
+
+		/*
+		 * No need to check the VM protection here.
+		 *
+		 * Maybe user space wants to query some of the data
+		 * when the VM is still unprotected. If we see the
+		 * need to fence a new data command we can still
+		 * return an error in the info handler.
+		 */
+
+		r = -EFAULT;
+		if (copy_from_user(&info, argp, sizeof(info.header)))
+			break;
+
+		r = -EINVAL;
+		if (info.header.len_max < sizeof(info.header))
+			break;
+
+		data_len = kvm_s390_handle_pv_info(&info);
+		if (data_len < 0) {
+			r = data_len;
+			break;
+		}
+		/*
+		 * If a data command struct is extended (multiple
+		 * times) this can be used to determine how much of it
+		 * is valid.
+		 */
+		info.header.len_written = data_len;
+
+		r = -EFAULT;
+		if (copy_to_user(argp, &info, data_len))
+			break;
+
+		r = 0;
+		break;
+	}
 	default:
 		r = -ENOTTY;
 	}
include/uapi/linux/kvm.h +25 -0
index 6a184d260c7f..8ac9de7d1386 100644
@@ -1653,6 +1653,30 @@ struct kvm_s390_pv_unp {
 	__u64 tweak;
 };
 
+enum pv_cmd_info_id {
+	KVM_PV_INFO_VM,
+};
+
+struct kvm_s390_pv_info_vm {
+	__u64 inst_calls_list[4];
+	__u64 max_cpus;
+	__u64 max_guests;
+	__u64 max_guest_addr;
+	__u64 feature_indication;
+};
+
+struct kvm_s390_pv_info_header {
+	__u32 id;
+	__u32 len_max;
+	__u32 len_written;
+	__u32 reserved;
+};
+
+struct kvm_s390_pv_info {
+	struct kvm_s390_pv_info_header header;
+	struct kvm_s390_pv_info_vm vm;
+};
+
 enum pv_cmd_id {
 	KVM_PV_ENABLE,
 	KVM_PV_DISABLE,
@@ -1661,6 +1685,7 @@ enum pv_cmd_id {
 	KVM_PV_VERIFY,
 	KVM_PV_PREP_RESET,
 	KVM_PV_UNSHARE_ALL,
+	KVM_PV_INFO,
 };
 
 struct kvm_pv_cmd {
Signature · 1 line
2.34.1
maintainer Reviewed-by on v5 3/10
#6 lore raw

Re: [PATCH v5 03/10] KVM: s390: pv: Add query interface

Quoted 4 lines · On Mon, 16 May 2022 09:08:10 +0000 Janosch Frank <[email protected]> wrote:

Some of the query information is already available via sysfs but having a IOCTL makes the information easier to retrieve.

Quoted 142 lines
---
 arch/s390/kvm/kvm-s390.c | 76 ++++++++++++++++++++++++++++++++++++++++
 include/uapi/linux/kvm.h | 25 +++++++++++++
 2 files changed, 101 insertions(+)
arch/s390/kvm/kvm-s390.c +76 -0
index 76ad6408cb2c..5859f243d287 100644
@@ -2224,6 +2224,42 @@ static int kvm_s390_cpus_to_pv(struct kvm *kvm, u16 *rc, u16 *rrc)
 	return r;
 }
 
+/*
+ * Here we provide user space with a direct interface to query UV
+ * related data like UV maxima and available features as well as
+ * feature specific data.
+ *
+ * To facilitate future extension of the data structures we'll try to
+ * write data up to the maximum requested length.
+ */
+static ssize_t kvm_s390_handle_pv_info(struct kvm_s390_pv_info *info)
+{
+	ssize_t len_min;
+
+	switch (info->header.id) {
+	case KVM_PV_INFO_VM: {
+		len_min =  sizeof(info->header) + sizeof(info->vm);
+
+		if (info->header.len_max < len_min)
+			return -EINVAL;
+
+		memcpy(info->vm.inst_calls_list,
+		       uv_info.inst_calls_list,
+		       sizeof(uv_info.inst_calls_list));
+
+		/* It's max cpuid not max cpus, so it's off by one */
+		info->vm.max_cpus = uv_info.max_guest_cpu_id + 1;
+		info->vm.max_guests = uv_info.max_num_sec_conf;
+		info->vm.max_guest_addr = uv_info.max_sec_stor_addr;
+		info->vm.feature_indication = uv_info.uv_feature_indications;
+
+		return len_min;
+	}
+	default:
+		return -EINVAL;
+	}
+}
+
 static int kvm_s390_handle_pv(struct kvm *kvm, struct kvm_pv_cmd *cmd)
 {
 	int r = 0;
@@ -2360,6 +2396,46 @@ static int kvm_s390_handle_pv(struct kvm *kvm, struct kvm_pv_cmd *cmd)
 			     cmd->rc, cmd->rrc);
 		break;
 	}
+	case KVM_PV_INFO: {
+		struct kvm_s390_pv_info info = {};
+		ssize_t data_len;
+
+		/*
+		 * No need to check the VM protection here.
+		 *
+		 * Maybe user space wants to query some of the data
+		 * when the VM is still unprotected. If we see the
+		 * need to fence a new data command we can still
+		 * return an error in the info handler.
+		 */
+
+		r = -EFAULT;
+		if (copy_from_user(&info, argp, sizeof(info.header)))
+			break;
+
+		r = -EINVAL;
+		if (info.header.len_max < sizeof(info.header))
+			break;
+
+		data_len = kvm_s390_handle_pv_info(&info);
+		if (data_len < 0) {
+			r = data_len;
+			break;
+		}
+		/*
+		 * If a data command struct is extended (multiple
+		 * times) this can be used to determine how much of it
+		 * is valid.
+		 */
+		info.header.len_written = data_len;
+
+		r = -EFAULT;
+		if (copy_to_user(argp, &info, data_len))
+			break;
+
+		r = 0;
+		break;
+	}
 	default:
 		r = -ENOTTY;
 	}
include/uapi/linux/kvm.h +25 -0
index 6a184d260c7f..8ac9de7d1386 100644
@@ -1653,6 +1653,30 @@ struct kvm_s390_pv_unp {
 	__u64 tweak;
 };
 
+enum pv_cmd_info_id {
+	KVM_PV_INFO_VM,
+};
+
+struct kvm_s390_pv_info_vm {
+	__u64 inst_calls_list[4];
+	__u64 max_cpus;
+	__u64 max_guests;
+	__u64 max_guest_addr;
+	__u64 feature_indication;
+};
+
+struct kvm_s390_pv_info_header {
+	__u32 id;
+	__u32 len_max;
+	__u32 len_written;
+	__u32 reserved;
+};
+
+struct kvm_s390_pv_info {
+	struct kvm_s390_pv_info_header header;
+	struct kvm_s390_pv_info_vm vm;
+};
+
 enum pv_cmd_id {
 	KVM_PV_ENABLE,
 	KVM_PV_DISABLE,
@@ -1661,6 +1685,7 @@ enum pv_cmd_id {
 	KVM_PV_VERIFY,
 	KVM_PV_PREP_RESET,
 	KVM_PV_UNSHARE_ALL,
+	KVM_PV_INFO,
 };
 
 struct kvm_pv_cmd {
author maintainer
#8 lore raw

PATCH v5 4/10 KVM: s390: pv: Add dump support definitions

In reply to Janosch Frank

Let's add the constants and structure definitions needed for the dump support.

---
 arch/s390/include/asm/uv.h | 33 +++++++++++++++++++++++++++++++++
 1 file changed, 33 insertions(+)
arch/s390/include/asm/uv.h +33 -0
index e8257a293dd1..3e597bb634bd 100644
@@ -50,6 +50,10 @@
 #define UVC_CMD_SET_UNSHARE_ALL		0x0340
 #define UVC_CMD_PIN_PAGE_SHARED		0x0341
 #define UVC_CMD_UNPIN_PAGE_SHARED	0x0342
+#define UVC_CMD_DUMP_INIT		0x0400
+#define UVC_CMD_DUMP_CONF_STOR_STATE	0x0401
+#define UVC_CMD_DUMP_CPU		0x0402
+#define UVC_CMD_DUMP_COMPLETE		0x0403
 #define UVC_CMD_SET_SHARED_ACCESS	0x1000
 #define UVC_CMD_REMOVE_SHARED_ACCESS	0x1001
 #define UVC_CMD_RETR_ATTEST		0x1020
@@ -77,6 +81,10 @@ enum uv_cmds_inst {
 	BIT_UVC_CMD_UNSHARE_ALL = 20,
 	BIT_UVC_CMD_PIN_PAGE_SHARED = 21,
 	BIT_UVC_CMD_UNPIN_PAGE_SHARED = 22,
+	BIT_UVC_CMD_DUMP_INIT = 24,
+	BIT_UVC_CMD_DUMP_CONFIG_STOR_STATE = 25,
+	BIT_UVC_CMD_DUMP_CPU = 26,
+	BIT_UVC_CMD_DUMP_COMPLETE = 27,
 	BIT_UVC_CMD_RETR_ATTEST = 28,
 };
 
@@ -246,6 +254,31 @@ struct uv_cb_attest {
 	u64 reserved168[4];		/* 0x0168 */
 } __packed __aligned(8);
 
+struct uv_cb_dump_cpu {
+	struct uv_cb_header header;
+	u64 reserved08[2];
+	u64 cpu_handle;
+	u64 dump_area_origin;
+	u64 reserved28[5];
+} __packed __aligned(8);
+
+struct uv_cb_dump_stor_state {
+	struct uv_cb_header header;
+	u64 reserved08[2];
+	u64 config_handle;
+	u64 dump_area_origin;
+	u64 gaddr;
+	u64 reserved28[4];
+} __packed __aligned(8);
+
+struct uv_cb_dump_complete {
+	struct uv_cb_header header;
+	u64 reserved08[2];
+	u64 config_handle;
+	u64 dump_area_origin;
+	u64 reserved30[5];
+} __packed __aligned(8);
+
 static inline int __uv_call(unsigned long r1, unsigned long r2)
 {
 	int cc;
Signature · 1 line
2.34.1
author maintainer
#10 lore raw

PATCH v5 5/10 KVM: s390: pv: Add query dump information

In reply to Janosch Frank

The dump API requires userspace to provide buffers into which we will store data. The dump information added in this patch tells userspace how big those buffers need to be.

---
 arch/s390/kvm/kvm-s390.c | 11 +++++++++++
 include/uapi/linux/kvm.h | 12 +++++++++++-
 2 files changed, 22 insertions(+), 1 deletion(-)
arch/s390/kvm/kvm-s390.c +11 -0
index 5859f243d287..de54f14e081e 100644
@@ -2255,6 +2255,17 @@ static ssize_t kvm_s390_handle_pv_info(struct kvm_s390_pv_info *info)
 
 		return len_min;
 	}
+	case KVM_PV_INFO_DUMP: {
+		len_min =  sizeof(info->header) + sizeof(info->dump);
+
+		if (info->header.len_max < len_min)
+			return -EINVAL;
+
+		info->dump.dump_cpu_buffer_len = uv_info.guest_cpu_stor_len;
+		info->dump.dump_config_mem_buffer_per_1m = uv_info.conf_dump_storage_state_len;
+		info->dump.dump_config_finalize_len = uv_info.conf_dump_finalize_len;
+		return len_min;
+	}
 	default:
 		return -EINVAL;
 	}
include/uapi/linux/kvm.h +11 -1
index 8ac9de7d1386..bb2f91bc2305 100644
@@ -1655,6 +1655,13 @@ struct kvm_s390_pv_unp {
 
 enum pv_cmd_info_id {
 	KVM_PV_INFO_VM,
+	KVM_PV_INFO_DUMP,
+};
+
+struct kvm_s390_pv_info_dump {
+	__u64 dump_cpu_buffer_len;
+	__u64 dump_config_mem_buffer_per_1m;
+	__u64 dump_config_finalize_len;
 };
 
 struct kvm_s390_pv_info_vm {
@@ -1674,7 +1681,10 @@ struct kvm_s390_pv_info_header {
 
 struct kvm_s390_pv_info {
 	struct kvm_s390_pv_info_header header;
-	struct kvm_s390_pv_info_vm vm;
+	union {
+		struct kvm_s390_pv_info_dump dump;
+		struct kvm_s390_pv_info_vm vm;
+	};
 };
 
 enum pv_cmd_id {
Signature · 1 line
2.34.1
maintainer Reviewed-by on v5 5/10
#11 lore raw

Re: [PATCH v5 05/10] KVM: s390: pv: Add query dump information

Quoted 5 lines · On Mon, 16 May 2022 09:08:12 +0000 Janosch Frank <[email protected]> wrote:

The dump API requires userspace to provide buffers into which we will store data. The dump information added in this patch tells userspace how big those buffers need to be.

Quoted 57 lines
---
 arch/s390/kvm/kvm-s390.c | 11 +++++++++++
 include/uapi/linux/kvm.h | 12 +++++++++++-
 2 files changed, 22 insertions(+), 1 deletion(-)
arch/s390/kvm/kvm-s390.c +11 -0
index 5859f243d287..de54f14e081e 100644
@@ -2255,6 +2255,17 @@ static ssize_t kvm_s390_handle_pv_info(struct kvm_s390_pv_info *info)
 
 		return len_min;
 	}
+	case KVM_PV_INFO_DUMP: {
+		len_min =  sizeof(info->header) + sizeof(info->dump);
+
+		if (info->header.len_max < len_min)
+			return -EINVAL;
+
+		info->dump.dump_cpu_buffer_len = uv_info.guest_cpu_stor_len;
+		info->dump.dump_config_mem_buffer_per_1m = uv_info.conf_dump_storage_state_len;
+		info->dump.dump_config_finalize_len = uv_info.conf_dump_finalize_len;
+		return len_min;
+	}
 	default:
 		return -EINVAL;
 	}
include/uapi/linux/kvm.h +11 -1
index 8ac9de7d1386..bb2f91bc2305 100644
@@ -1655,6 +1655,13 @@ struct kvm_s390_pv_unp {
 
 enum pv_cmd_info_id {
 	KVM_PV_INFO_VM,
+	KVM_PV_INFO_DUMP,
+};
+
+struct kvm_s390_pv_info_dump {
+	__u64 dump_cpu_buffer_len;
+	__u64 dump_config_mem_buffer_per_1m;
+	__u64 dump_config_finalize_len;
 };
 
 struct kvm_s390_pv_info_vm {
@@ -1674,7 +1681,10 @@ struct kvm_s390_pv_info_header {
 
 struct kvm_s390_pv_info {
 	struct kvm_s390_pv_info_header header;
-	struct kvm_s390_pv_info_vm vm;
+	union {
+		struct kvm_s390_pv_info_dump dump;
+		struct kvm_s390_pv_info_vm vm;
+	};
 };
 
 enum pv_cmd_id {
author maintainer
#13 lore raw

PATCH v5 6/10 kvm: s390: Add configuration dump functionality

In reply to Janosch Frank

Sometimes dumping inside of a VM fails, is unavailable or doesn't yield the required data. For these occasions we dump the VM from the outside, writing memory and cpu data to a file.

Up to now PV guests only supported dumping from the inside of the guest through dumpers like KDUMP. A PV guest can be dumped from the hypervisor but the data will be stale and / or encrypted.

To get the actual state of the PV VM we need the help of the Ultravisor who safeguards the VM state. New UV calls have been added to initialize the dump, dump storage state data, dump cpu data and complete the dump process. We expose these calls in this patch via a new UV ioctl command.

The sensitive parts of the dump data are encrypted, the dump key is derived from the Customer Communication Key (CCK). This ensures that only the owner of the VM who has the CCK can decrypt the dump data.

The memory is dumped / read via a normal export call and a re-import after the dump initialization is not needed (no re-encryption with a dump key).

---
 arch/s390/include/asm/kvm_host.h |   1 +
 arch/s390/kvm/kvm-s390.c         | 126 +++++++++++++++++++++++++++++++
 arch/s390/kvm/kvm-s390.h         |   2 +
 arch/s390/kvm/pv.c               | 113 +++++++++++++++++++++++++++
 include/uapi/linux/kvm.h         |  15 ++++
 5 files changed, 257 insertions(+)
arch/s390/include/asm/kvm_host.h +1 -0
index 766028d54a3e..a0fbe4820e0a 100644
@@ -923,6 +923,7 @@ struct kvm_s390_pv {
 	u64 guest_len;
 	unsigned long stor_base;
 	void *stor_var;
+	bool dumping;
 };
 
 struct kvm_arch{
arch/s390/kvm/kvm-s390.c +126 -0

144 lines not on this page · raw message

arch/s390/kvm/kvm-s390.h +2 -0
index 497d52a83c78..2868dd0bba25 100644
@@ -250,6 +250,8 @@ int kvm_s390_pv_set_sec_parms(struct kvm *kvm, void *hdr, u64 length, u16 *rc,
 int kvm_s390_pv_unpack(struct kvm *kvm, unsigned long addr, unsigned long size,
 		       unsigned long tweak, u16 *rc, u16 *rrc);
 int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state);
+int kvm_s390_pv_dump_stor_state(struct kvm *kvm, void __user *buff_user,
+				u64 *gaddr, u64 buff_user_len, u16 *rc, u16 *rrc);
 
 static inline u64 kvm_s390_pv_get_handle(struct kvm *kvm)
 {
arch/s390/kvm/pv.c +113 -0

122 lines not on this page · raw message

include/uapi/linux/kvm.h +15 -0
index bb2f91bc2305..1c60c2d314ba 100644
@@ -1653,6 +1653,20 @@ struct kvm_s390_pv_unp {
 	__u64 tweak;
 };
 
+enum pv_cmd_dmp_id {
+	KVM_PV_DUMP_INIT,
+	KVM_PV_DUMP_CONFIG_STOR_STATE,
+	KVM_PV_DUMP_COMPLETE,
+};
+
+struct kvm_s390_pv_dmp {
+	__u64 subcmd;
+	__u64 buff_addr;
+	__u64 buff_len;
+	__u64 gaddr;		/* For dump storage state */
+	__u64 reserved[4];
+};
+
 enum pv_cmd_info_id {
 	KVM_PV_INFO_VM,
 	KVM_PV_INFO_DUMP,
@@ -1696,6 +1710,7 @@ enum pv_cmd_id {
 	KVM_PV_PREP_RESET,
 	KVM_PV_UNSHARE_ALL,
 	KVM_PV_INFO,
+	KVM_PV_DUMP,
 };
 
 struct kvm_pv_cmd {
Signature · 1 line
2.34.1
maintainer
#14 lore raw

Re: [PATCH v5 06/10] kvm: s390: Add configuration dump functionality

Quoted 55 lines · On Mon, 16 May 2022 09:08:13 +0000 Janosch Frank <[email protected]> wrote:

Sometimes dumping inside of a VM fails, is unavailable or doesn't yield the required data. For these occasions we dump the VM from the outside, writing memory and cpu data to a file.

Up to now PV guests only supported dumping from the inside of the guest through dumpers like KDUMP. A PV guest can be dumped from the hypervisor but the data will be stale and / or encrypted.

To get the actual state of the PV VM we need the help of the Ultravisor who safeguards the VM state. New UV calls have been added to initialize the dump, dump storage state data, dump cpu data and complete the dump process. We expose these calls in this patch via a new UV ioctl command.

The sensitive parts of the dump data are encrypted, the dump key is derived from the Customer Communication Key (CCK). This ensures that only the owner of the VM who has the CCK can decrypt the dump data.

The memory is dumped / read via a normal export call and a re-import after the dump initialization is not needed (no re-encryption with a dump key).

---
 arch/s390/include/asm/kvm_host.h |   1 +
 arch/s390/kvm/kvm-s390.c         | 126 +++++++++++++++++++++++++++++++
 arch/s390/kvm/kvm-s390.h         |   2 +
 arch/s390/kvm/pv.c               | 113 +++++++++++++++++++++++++++
 include/uapi/linux/kvm.h         |  15 ++++
 5 files changed, 257 insertions(+)
arch/s390/include/asm/kvm_host.h +1 -0
index 766028d54a3e..a0fbe4820e0a 100644
@@ -923,6 +923,7 @@ struct kvm_s390_pv {
 	u64 guest_len;
 	unsigned long stor_base;
 	void *stor_var;
+	bool dumping;
 };
 
 struct kvm_arch{
arch/s390/kvm/kvm-s390.c +4 -0
index de54f14e081e..6bf9dd85d50f 100644
@@ -2271,6 +2271,101 @@ static ssize_t kvm_s390_handle_pv_info(struct kvm_s390_pv_info *info)
 	}
 }
 
+static int kvm_s390_pv_dmp(struct kvm *kvm, struct kvm_pv_cmd *cmd,
+			   struct kvm_s390_pv_dmp dmp)
+{
+	int r = -EINVAL;

arch/s390/kvm/kvm-s390.c static ssize_t kvm_s390_handle_pv_info(struct kvm_s390_pv_info *info) · line 2287

Quoted 10 lines
arch/s390/kvm/kvm-s390.c +10 -0
+	void __user *result_buff = (void __user *)dmp.buff_addr;
+
+	switch (dmp.subcmd) {
+	case KVM_PV_DUMP_INIT: {
+		if (kvm->arch.pv.dumping)
+			break;
+
+		/*
+		 * Block SIE entry as concurrent dump UVCs could lead
+		 * to validities.

to fatal validity intercepts

Quoted 19 lines
arch/s390/kvm/kvm-s390.c +19 -0
+		 */
+		kvm_s390_vcpu_block_all(kvm);
+
+		r = uv_cmd_nodata(kvm_s390_pv_get_handle(kvm),
+				  UVC_CMD_DUMP_INIT, &cmd->rc, &cmd->rrc);
+		KVM_UV_EVENT(kvm, 3, "PROTVIRT DUMP INIT: rc %x rrc %x",
+			     cmd->rc, cmd->rrc);
+		if (!r) {
+			kvm->arch.pv.dumping = true;
+		} else {
+			kvm_s390_vcpu_unblock_all(kvm);
+			r = -EINVAL;
+		}
+		break;
+	}
+	case KVM_PV_DUMP_CONFIG_STOR_STATE: {
+		if (!kvm->arch.pv.dumping)
+			break;
+

arch/s390/kvm/kvm-s390.c static ssize_t kvm_s390_handle_pv_info(struct kvm_s390_pv_info *info) · line 2316

Quoted 10 lines
arch/s390/kvm/kvm-s390.c +10 -0
+		/*
+		 * gaddr is an output parameter since we might stop
+		 * early. As dmp will be copied back in our caller, we
+		 * don't need to do it ourselves.
+		 */
+		r = kvm_s390_pv_dump_stor_state(kvm, result_buff, &dmp.gaddr, dmp.buff_len,
+						&cmd->rc, &cmd->rrc);
+		break;
+	}
+	case KVM_PV_DUMP_COMPLETE: {

this starts to be quite complex, maybe push it into a separate function as well, like you did for KVM_PV_DUMP_CONFIG_STOR_STATE

Quoted 85 lines
arch/s390/kvm/kvm-s390.c +74 -0
+		struct uv_cb_dump_complete complete = {
+			.header.len = sizeof(complete),
+			.header.cmd = UVC_CMD_DUMP_COMPLETE,
+			.config_handle = kvm_s390_pv_get_handle(kvm),
+		};
+		u64 *compl_data;
+
+		r = -EINVAL;
+		if (!kvm->arch.pv.dumping)
+			break;
+
+		if (dmp.buff_len < uv_info.conf_dump_finalize_len)
+			break;
+
+		/* Allocate dump area */
+		r = -ENOMEM;
+		compl_data = vzalloc(uv_info.conf_dump_finalize_len);
+		if (!compl_data)
+			break;
+		complete.dump_area_origin = (u64)compl_data;
+
+		r = uv_call_sched(0, (u64)&complete);
+		cmd->rc = complete.header.rc;
+		cmd->rrc = complete.header.rrc;
+		KVM_UV_EVENT(kvm, 3, "PROTVIRT DUMP COMPLETE: rc %x rrc %x",
+			     complete.header.rc, complete.header.rrc);
+
+		if (!r) {
+			/*
+			 * kvm_s390_pv_dealloc_vm() will also (mem)set
+			 * this to false on a reboot or other destroy
+			 * operation for this vm.
+			 */
+			kvm->arch.pv.dumping = false;
+			kvm_s390_vcpu_unblock_all(kvm);
+			r = copy_to_user(result_buff, compl_data, uv_info.conf_dump_finalize_len);
+			if (r)
+				r = -EFAULT;
+		}
+		vfree(compl_data);
+		if (r > 0)
+			r = -EINVAL;
+		break;
+	}
+	default:
+		r = -ENOTTY;
+		break;
+	}
+
+	return r;
+}
+
 static int kvm_s390_handle_pv(struct kvm *kvm, struct kvm_pv_cmd *cmd)
 {
 	int r = 0;
@@ -2447,6 +2542,28 @@ static int kvm_s390_handle_pv(struct kvm *kvm, struct kvm_pv_cmd *cmd)
 		r = 0;
 		break;
 	}
+	case KVM_PV_DUMP: {
+		struct kvm_s390_pv_dmp dmp;
+
+		r = -EINVAL;
+		if (!kvm_s390_pv_is_protected(kvm))
+			break;
+
+		r = -EFAULT;
+		if (copy_from_user(&dmp, argp, sizeof(dmp)))
+			break;
+
+		r = kvm_s390_pv_dmp(kvm, cmd, dmp);
+		if (r)
+			break;
+
+		if (copy_to_user(argp, &dmp, sizeof(dmp))) {
+			r = -EFAULT;
+			break;
+		}
+
+		break;
+	}
 	default:
 		r = -ENOTTY;
 	}

arch/s390/kvm/kvm-s390.c int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu) · line 4687

Quoted 7 lines
arch/s390/kvm/kvm-s390.c +4 -0
@@ -4564,6 +4681,15 @@ int kvm_arch_vcpu_ioctl_run(struct kvm_vcpu *vcpu)
 	struct kvm_run *kvm_run = vcpu->run;
 	int rc;
 
+	/*
+	 * Running a VM while dumping always has the potential to
+	 * produce inconsistent dump data. But for PV vcpus a SIE
+	 * entry while dumping could also lead to a validity which we

"a fatal validity intercept"

Quoted 44 lines
arch/s390/kvm/kvm-s390.c +5 -0
+	 * absolutely want to avoid.
+	 */
+	if (vcpu->kvm->arch.pv.dumping)
+		return -EINVAL;
+
 	if (kvm_run->immediate_exit)
 		return -EINTR;
 
arch/s390/kvm/kvm-s390.h +2 -0
index 497d52a83c78..2868dd0bba25 100644
@@ -250,6 +250,8 @@ int kvm_s390_pv_set_sec_parms(struct kvm *kvm, void *hdr, u64 length, u16 *rc,
 int kvm_s390_pv_unpack(struct kvm *kvm, unsigned long addr, unsigned long size,
 		       unsigned long tweak, u16 *rc, u16 *rrc);
 int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state);
+int kvm_s390_pv_dump_stor_state(struct kvm *kvm, void __user *buff_user,
+				u64 *gaddr, u64 buff_user_len, u16 *rc, u16 *rrc);
 
 static inline u64 kvm_s390_pv_get_handle(struct kvm *kvm)
 {
arch/s390/kvm/pv.c +8 -0
index cc7c9599f43e..fd261667d2c2 100644
@@ -7,6 +7,7 @@
  */
 #include <linux/kvm.h>
 #include <linux/kvm_host.h>
+#include <linux/minmax.h>
 #include <linux/pagemap.h>
 #include <linux/sched/signal.h>
 #include <asm/gmap.h>
@@ -298,3 +299,115 @@ int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state)
 		return -EINVAL;
 	return 0;
 }
+
+/* Size of the cache for the storage state dump data. 1MB for now */
+#define DUMP_BUFF_LEN HPAGE_SIZE
+
+/*
+ * kvm_s390_pv_dump_stor_state
+ *

arch/s390/kvm/pv.c int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state) · line 318

Quoted 10 lines
arch/s390/kvm/pv.c +10 -0
+ * @kvm: pointer to the guest's KVM struct
+ * @buff_user: Userspace pointer where we will write the results to
+ * @gaddr: Starting absolute guest address for which the storage state
+ *         is requested. This value will be updated with the last
+ *         address for which data was written when returning to
+ *         userspace.
+ * @buff_user_len: Length of the buff_user buffer
+ * @rc: Pointer to where the uvcb return code is stored
+ * @rrc: Pointer to where the uvcb return reason code is stored
+ *

please add:

	Context: kvm->lock needs to be held

also explain that part of the user buffer might be written to even in case of failure (this also needs to go in the documentation)

arch/s390/kvm/pv.c int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state) · line 320

Quoted 2 lines
arch/s390/kvm/pv.c +2 -0
+ * Return:
+ *  0 on success

rc and rrc will also be set in case of success

arch/s390/kvm/pv.c int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state) · line 324

Quoted 4 lines
arch/s390/kvm/pv.c +4 -0
+ *  -ENOMEM if allocating the cache fails
+ *  -EINVAL if gaddr is not aligned to 1MB
+ *  -EINVAL if buff_user_len is not aligned to uv_info.conf_dump_storage_state_len
+ *  -EINVAL if the UV call fails, rc and rrc will be set in this case

have you considered a different code for UVC failure? so the caller can know that rc and rrc are meaningful

or just explain that rc and rrc will always be set; if the UVC is not performed, rc and rrc will be 0

Quoted 47 lines
arch/s390/kvm/pv.c +47 -0
+ *  -EFAULT if copying the result to buff_user failed
+ */
+int kvm_s390_pv_dump_stor_state(struct kvm *kvm, void __user *buff_user,
+				u64 *gaddr, u64 buff_user_len, u16 *rc, u16 *rrc)
+{
+	struct uv_cb_dump_stor_state uvcb = {
+		.header.cmd = UVC_CMD_DUMP_CONF_STOR_STATE,
+		.header.len = sizeof(uvcb),
+		.config_handle = kvm->arch.pv.handle,
+		.gaddr = *gaddr,
+		.dump_area_origin = 0,
+	};
+	size_t buff_kvm_size;
+	size_t size_done = 0;
+	u8 *buff_kvm = NULL;
+	int cc, ret;
+
+	ret = -EINVAL;
+	/* UV call processes 1MB guest storage chunks at a time */
+	if (!IS_ALIGNED(*gaddr, HPAGE_SIZE))
+		goto out;
+
+	/*
+	 * We provide the storage state for 1MB chunks of guest
+	 * storage. The buffer will need to be aligned to
+	 * conf_dump_storage_state_len so we don't end on a partial
+	 * chunk.
+	 */
+	if (!buff_user_len ||
+	    !IS_ALIGNED(buff_user_len, uv_info.conf_dump_storage_state_len))
+		goto out;
+
+	/*
+	 * Allocate a buffer from which we will later copy to the user
+	 * process. We don't want userspace to dictate our buffer size
+	 * so we limit it to DUMP_BUFF_LEN.
+	 */
+	ret = -ENOMEM;
+	buff_kvm_size = min_t(u64, buff_user_len, DUMP_BUFF_LEN);
+	buff_kvm = vzalloc(buff_kvm_size);
+	if (!buff_kvm)
+		goto out;
+
+	ret = 0;
+	uvcb.dump_area_origin = (u64)buff_kvm;
+	/* We will loop until the user buffer is filled or an error occurs */
+	do {

arch/s390/kvm/pv.c int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state) · line 381

Quoted 10 lines
arch/s390/kvm/pv.c +10 -0
+		/* Get 1MB worth of guest storage state data */
+		cc = uv_call_sched(0, (u64)&uvcb);
+
+		/* All or nothing */
+		if (cc) {
+			ret = -EINVAL;
+			break;
+		}
+
+		size_done += uv_info.conf_dump_storage_state_len;

maybe save this in a local const variable with a shorter name? would be more readable? const u64 dump_len = uv_info.conf_dump_storage_state_len;

Quoted 65 lines
arch/s390/kvm/pv.c +32 -0
+		uvcb.dump_area_origin += uv_info.conf_dump_storage_state_len;
+		buff_user_len -= uv_info.conf_dump_storage_state_len;
+		uvcb.gaddr += HPAGE_SIZE;
+
+		/* KVM Buffer full, time to copy to the process */
+		if (!buff_user_len || size_done == DUMP_BUFF_LEN) {
+			if (copy_to_user(buff_user, buff_kvm, size_done)) {
+				ret = -EFAULT;
+				break;
+			}
+
+			buff_user += size_done;
+			size_done = 0;
+			uvcb.dump_area_origin = (u64)buff_kvm;
+		}
+	} while (buff_user_len);
+
+	/* Report back where we ended dumping */
+	*gaddr = uvcb.gaddr;
+
+	/* Lets only log errors, we don't want to spam */
+out:
+	if (ret)
+		KVM_UV_EVENT(kvm, 3,
+			     "PROTVIRT DUMP STORAGE STATE: addr %llx ret %d, uvcb rc %x rrc %x",
+			     uvcb.gaddr, ret, uvcb.header.rc, uvcb.header.rrc);
+	*rc = uvcb.header.rc;
+	*rrc = uvcb.header.rrc;
+	vfree(buff_kvm);
+
+	return ret;
+}
include/uapi/linux/kvm.h +15 -0
index bb2f91bc2305..1c60c2d314ba 100644
@@ -1653,6 +1653,20 @@ struct kvm_s390_pv_unp {
 	__u64 tweak;
 };
 
+enum pv_cmd_dmp_id {
+	KVM_PV_DUMP_INIT,
+	KVM_PV_DUMP_CONFIG_STOR_STATE,
+	KVM_PV_DUMP_COMPLETE,
+};
+
+struct kvm_s390_pv_dmp {
+	__u64 subcmd;
+	__u64 buff_addr;
+	__u64 buff_len;
+	__u64 gaddr;		/* For dump storage state */
+	__u64 reserved[4];
+};
+
 enum pv_cmd_info_id {
 	KVM_PV_INFO_VM,
 	KVM_PV_INFO_DUMP,
@@ -1696,6 +1710,7 @@ enum pv_cmd_id {
 	KVM_PV_PREP_RESET,
 	KVM_PV_UNSHARE_ALL,
 	KVM_PV_INFO,
+	KVM_PV_DUMP,
 };
 
 struct kvm_pv_cmd {
author maintainer
#15 lore raw

Re: [PATCH v5 06/10] kvm: s390: Add configuration dump functionality

Quoted 21 lines · Claudio Imbrenda
Quoted 18 lines · On Mon, 16 May 2022 09:08:13 +0000 Janosch Frank <[email protected]> wrote:

Sometimes dumping inside of a VM fails, is unavailable or doesn't yield the required data. For these occasions we dump the VM from the outside, writing memory and cpu data to a file.

Up to now PV guests only supported dumping from the inside of the guest through dumpers like KDUMP. A PV guest can be dumped from the hypervisor but the data will be stale and / or encrypted.

To get the actual state of the PV VM we need the help of the Ultravisor who safeguards the VM state. New UV calls have been added to initialize the dump, dump storage state data, dump cpu data and complete the dump process. We expose these calls in this patch via a new UV ioctl command.

The sensitive parts of the dump data are encrypted, the dump key is derived from the Customer Communication Key (CCK). This ensures that only the owner of the VM who has the CCK can decrypt the dump data.

Quoted 5 lines
Quoted 5 lines

The memory is dumped / read via a normal export call and a re-import after the dump initialization is not needed (no re-encryption with a dump key).

The cut code will be fixed according to your requests.

[...]

Quoted 14 lines
Quoted 6 lines

+/*
+ * kvm_s390_pv_dump_stor_state
+ *
+ * @kvm: pointer to the guest's KVM struct
+ * @buff_user: Userspace pointer where we will write the results to
+ * @gaddr: Starting absolute guest address for which the storage state

Quoted 7 lines
+ *         is requested. This value will be updated with the last
+ *         address for which data was written when returning to
+ *         userspace.

+ * @buff_user_len: Length of the buff_user buffer
+ * @rc: Pointer to where the uvcb return code is stored
+ * @rrc: Pointer to where the uvcb return reason code is stored
+ *

Quoted 2 lines

please add:

	Context: kvm->lock needs to be held

Sure

Quoted 3 lines

also explain that part of the user buffer might be written to even in case of failure (this also needs to go in the documentation)

Ok

Quoted 5 lines
Quoted 3 lines

+ * Return:
+ * 0 on success

rc and rrc will also be set in case of success

But that's different from the return code of this function and would belong to the function description above, no?

Quoted 11 lines
Quoted 5 lines
+ *  -ENOMEM if allocating the cache fails
+ *  -EINVAL if gaddr is not aligned to 1MB
+ *  -EINVAL if buff_user_len is not aligned to uv_info.conf_dump_storage_state_len
+ *  -EINVAL if the UV call fails, rc and rrc will be set in this case

have you considered a different code for UVC failure? so the caller can know that rc and rrc are meaningful

or just explain that rc and rrc will always be set; if the UVC is not performed, rc and rrc will be 0

If the UVC is not performed the rcs will not be *changed*, so it's advisable to set them to 0 to recognize a change.

Also:
While I generally like commenting as much as possible, this is starting to get out of hand, the comment header is now taking up a lot of space. I'll put the rc/rrc comment into the api documentation and I'm considering putting the partial write comment into there too.

Quoted 128 lines
Quoted 48 lines

+ * -EFAULT if copying the result to buff_user failed
+ */

arch/s390/kvm/pv.c +45 -0
+int kvm_s390_pv_dump_stor_state(struct kvm *kvm, void __user *buff_user,
+				u64 *gaddr, u64 buff_user_len, u16 *rc, u16 *rrc)
+{
+	struct uv_cb_dump_stor_state uvcb = {
+		.header.cmd = UVC_CMD_DUMP_CONF_STOR_STATE,
+		.header.len = sizeof(uvcb),
+		.config_handle = kvm->arch.pv.handle,
+		.gaddr = *gaddr,
+		.dump_area_origin = 0,
+	};
+	size_t buff_kvm_size;
+	size_t size_done = 0;
+	u8 *buff_kvm = NULL;
+	int cc, ret;
+
+	ret = -EINVAL;
+	/* UV call processes 1MB guest storage chunks at a time */
+	if (!IS_ALIGNED(*gaddr, HPAGE_SIZE))
+		goto out;
+
+	/*
+	 * We provide the storage state for 1MB chunks of guest
+	 * storage. The buffer will need to be aligned to
+	 * conf_dump_storage_state_len so we don't end on a partial
+	 * chunk.
+	 */
+	if (!buff_user_len ||
+	    !IS_ALIGNED(buff_user_len, uv_info.conf_dump_storage_state_len))
+		goto out;
+
+	/*
+	 * Allocate a buffer from which we will later copy to the user
+	 * process. We don't want userspace to dictate our buffer size
+	 * so we limit it to DUMP_BUFF_LEN.
+	 */
+	ret = -ENOMEM;
+	buff_kvm_size = min_t(u64, buff_user_len, DUMP_BUFF_LEN);
+	buff_kvm = vzalloc(buff_kvm_size);
+	if (!buff_kvm)
+		goto out;
+
+	ret = 0;
+	uvcb.dump_area_origin = (u64)buff_kvm;
+	/* We will loop until the user buffer is filled or an error occurs */
+	do {

arch/s390/kvm/pv.c int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state) · line 381

Quoted 10 lines
arch/s390/kvm/pv.c +10 -0
+		/* Get 1MB worth of guest storage state data */
+		cc = uv_call_sched(0, (u64)&uvcb);
+
+		/* All or nothing */
+		if (cc) {
+			ret = -EINVAL;
+			break;
+		}
+
+		size_done += uv_info.conf_dump_storage_state_len;

maybe save this in a local const variable with a shorter name? would be more readable? const u64 dump_len = uv_info.conf_dump_storage_state_len;

Quoted 65 lines
arch/s390/kvm/pv.c +32 -0
+		uvcb.dump_area_origin += uv_info.conf_dump_storage_state_len;
+		buff_user_len -= uv_info.conf_dump_storage_state_len;
+		uvcb.gaddr += HPAGE_SIZE;
+
+		/* KVM Buffer full, time to copy to the process */
+		if (!buff_user_len || size_done == DUMP_BUFF_LEN) {
+			if (copy_to_user(buff_user, buff_kvm, size_done)) {
+				ret = -EFAULT;
+				break;
+			}
+
+			buff_user += size_done;
+			size_done = 0;
+			uvcb.dump_area_origin = (u64)buff_kvm;
+		}
+	} while (buff_user_len);
+
+	/* Report back where we ended dumping */
+	*gaddr = uvcb.gaddr;
+
+	/* Lets only log errors, we don't want to spam */
+out:
+	if (ret)
+		KVM_UV_EVENT(kvm, 3,
+			     "PROTVIRT DUMP STORAGE STATE: addr %llx ret %d, uvcb rc %x rrc %x",
+			     uvcb.gaddr, ret, uvcb.header.rc, uvcb.header.rrc);
+	*rc = uvcb.header.rc;
+	*rrc = uvcb.header.rrc;
+	vfree(buff_kvm);
+
+	return ret;
+}
include/uapi/linux/kvm.h +15 -0
index bb2f91bc2305..1c60c2d314ba 100644
@@ -1653,6 +1653,20 @@ struct kvm_s390_pv_unp {
  	__u64 tweak;
  };
 
+enum pv_cmd_dmp_id {
+	KVM_PV_DUMP_INIT,
+	KVM_PV_DUMP_CONFIG_STOR_STATE,
+	KVM_PV_DUMP_COMPLETE,
+};
+
+struct kvm_s390_pv_dmp {
+	__u64 subcmd;
+	__u64 buff_addr;
+	__u64 buff_len;
+	__u64 gaddr;		/* For dump storage state */
+	__u64 reserved[4];
+};
+
  enum pv_cmd_info_id {
  	KVM_PV_INFO_VM,
  	KVM_PV_INFO_DUMP,
@@ -1696,6 +1710,7 @@ enum pv_cmd_id {
  	KVM_PV_PREP_RESET,
  	KVM_PV_UNSHARE_ALL,
  	KVM_PV_INFO,
+	KVM_PV_DUMP,
  };
 
  struct kvm_pv_cmd {
maintainer
#16 lore raw

Re: [PATCH v5 06/10] kvm: s390: Add configuration dump functionality

Quoted 26 lines · On Tue, 17 May 2022 15:39:14 +0200 Janosch Frank <[email protected]> wrote:
Quoted 25 lines · Claudio Imbrenda
Quoted 22 lines · On Mon, 16 May 2022 09:08:13 +0000 Janosch Frank <[email protected]> wrote:

Sometimes dumping inside of a VM fails, is unavailable or doesn't yield the required data. For these occasions we dump the VM from the outside, writing memory and cpu data to a file.

Up to now PV guests only supported dumping from the inside of the guest through dumpers like KDUMP. A PV guest can be dumped from the hypervisor but the data will be stale and / or encrypted.

To get the actual state of the PV VM we need the help of the Ultravisor who safeguards the VM state. New UV calls have been added to initialize the dump, dump storage state data, dump cpu data and complete the dump process. We expose these calls in this patch via a new UV ioctl command.

The sensitive parts of the dump data are encrypted, the dump key is derived from the Customer Communication Key (CCK). This ensures that only the owner of the VM who has the CCK can decrypt the dump data.

The memory is dumped / read via a normal export call and a re-import after the dump initialization is not needed (no re-encryption with a dump key).

Quoted 6 lines
Quoted 1 line
Quoted 1 line

The cut code will be fixed according to your requests.

[...]

Quoted 1 line
Quoted 1 line

+/*

/**

Quoted 25 lines
Quoted 13 lines
Quoted 5 lines

+ * kvm_s390_pv_dump_stor_state
+ *
+ * @kvm: pointer to the guest's KVM struct
+ * @buff_user: Userspace pointer where we will write the results to
+ * @gaddr: Starting absolute guest address for which the storage state

Quoted 7 lines
+ *         is requested. This value will be updated with the last
+ *         address for which data was written when returning to
+ *         userspace.

+ * @buff_user_len: Length of the buff_user buffer
+ * @rc: Pointer to where the uvcb return code is stored
+ * @rrc: Pointer to where the uvcb return reason code is stored
+ *

Quoted 2 lines

please add:

	Context: kvm->lock needs to be held  

Sure

Quoted 3 lines

also explain that part of the user buffer might be written to even in case of failure (this also needs to go in the documentation)

Ok

Quoted 1 line
Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

+ * Return:
+ * 0 on success

rc and rrc will also be set in case of success

But that's different from the return code of this function and would belong to the function description above, no?

yes

Quoted 7 lines
Quoted 7 lines
Quoted 6 lines
  
+ *  -ENOMEM if allocating the cache fails
+ *  -EINVAL if gaddr is not aligned to 1MB
+ *  -EINVAL if buff_user_len is not aligned to uv_info.conf_dump_storage_state_len
+ *  -EINVAL if the UV call fails, rc and rrc will be set in this case  
Quoted 8 lines
Quoted 5 lines

have you considered a different code for UVC failure? so the caller can know that rc and rrc are meaningful

or just explain that rc and rrc will always be set; if the UVC is not performed, rc and rrc will be 0

If the UVC is not performed the rcs will not be *changed*, so it's advisable to set them to 0 to recognize a change.

ah, right, you jump to out only for errors after the allocation

Quoted 7 lines

Also:
While I generally like commenting as much as possible, this is starting to get out of hand, the comment header is now taking up a lot of space. I'll put the rc/rrc comment into the api documentation and I'm considering putting the partial write comment into there too.

I don't know, the comments are also used to generate documentation automatically, so I think they should contain all the information in one way or the other.

I don't see issues with huge comments, if the function is complex and has lots of parameters and possible return values.

as much as it bothers you, the whole comment block has 0 lines of actual description of what the function does, so I don't think this is getting out of hand :)

Quoted 55 lines
Quoted 55 lines
Quoted 55 lines

+ * -EFAULT if copying the result to buff_user failed
+ */

arch/s390/kvm/pv.c +51 -0
+int kvm_s390_pv_dump_stor_state(struct kvm *kvm, void __user *buff_user,
+				u64 *gaddr, u64 buff_user_len, u16 *rc, u16 *rrc)
+{
+	struct uv_cb_dump_stor_state uvcb = {
+		.header.cmd = UVC_CMD_DUMP_CONF_STOR_STATE,
+		.header.len = sizeof(uvcb),
+		.config_handle = kvm->arch.pv.handle,
+		.gaddr = *gaddr,
+		.dump_area_origin = 0,
+	};
+	size_t buff_kvm_size;
+	size_t size_done = 0;
+	u8 *buff_kvm = NULL;
+	int cc, ret;
+
+	ret = -EINVAL;
+	/* UV call processes 1MB guest storage chunks at a time */
+	if (!IS_ALIGNED(*gaddr, HPAGE_SIZE))
+		goto out;
+
+	/*
+	 * We provide the storage state for 1MB chunks of guest
+	 * storage. The buffer will need to be aligned to
+	 * conf_dump_storage_state_len so we don't end on a partial
+	 * chunk.
+	 */
+	if (!buff_user_len ||
+	    !IS_ALIGNED(buff_user_len, uv_info.conf_dump_storage_state_len))
+		goto out;
+
+	/*
+	 * Allocate a buffer from which we will later copy to the user
+	 * process. We don't want userspace to dictate our buffer size
+	 * so we limit it to DUMP_BUFF_LEN.
+	 */
+	ret = -ENOMEM;
+	buff_kvm_size = min_t(u64, buff_user_len, DUMP_BUFF_LEN);
+	buff_kvm = vzalloc(buff_kvm_size);
+	if (!buff_kvm)
+		goto out;
+
+	ret = 0;
+	uvcb.dump_area_origin = (u64)buff_kvm;
+	/* We will loop until the user buffer is filled or an error occurs */
+	do {
+		/* Get 1MB worth of guest storage state data */
+		cc = uv_call_sched(0, (u64)&uvcb);
+
+		/* All or nothing */
+		if (cc) {
+			ret = -EINVAL;
Quoted 8 lines
Quoted 8 lines

arch/s390/kvm/pv.c int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state) · line 381

Quoted 4 lines
arch/s390/kvm/pv.c +4 -0
+			break;
+		}
+
+		size_done += uv_info.conf_dump_storage_state_len;  

maybe save this in a local const variable with a shorter name? would be more readable? const u64 dump_len = uv_info.conf_dump_storage_state_len;

did you see this one ^ ?

Quoted 67 lines
Quoted 66 lines
Quoted 65 lines
arch/s390/kvm/pv.c +32 -0
+		uvcb.dump_area_origin += uv_info.conf_dump_storage_state_len;
+		buff_user_len -= uv_info.conf_dump_storage_state_len;
+		uvcb.gaddr += HPAGE_SIZE;
+
+		/* KVM Buffer full, time to copy to the process */
+		if (!buff_user_len || size_done == DUMP_BUFF_LEN) {
+			if (copy_to_user(buff_user, buff_kvm, size_done)) {
+				ret = -EFAULT;
+				break;
+			}
+
+			buff_user += size_done;
+			size_done = 0;
+			uvcb.dump_area_origin = (u64)buff_kvm;
+		}
+	} while (buff_user_len);
+
+	/* Report back where we ended dumping */
+	*gaddr = uvcb.gaddr;
+
+	/* Lets only log errors, we don't want to spam */
+out:
+	if (ret)
+		KVM_UV_EVENT(kvm, 3,
+			     "PROTVIRT DUMP STORAGE STATE: addr %llx ret %d, uvcb rc %x rrc %x",
+			     uvcb.gaddr, ret, uvcb.header.rc, uvcb.header.rrc);
+	*rc = uvcb.header.rc;
+	*rrc = uvcb.header.rrc;
+	vfree(buff_kvm);
+
+	return ret;
+}
include/uapi/linux/kvm.h +15 -0
index bb2f91bc2305..1c60c2d314ba 100644
@@ -1653,6 +1653,20 @@ struct kvm_s390_pv_unp {
  	__u64 tweak;
  };
 
+enum pv_cmd_dmp_id {
+	KVM_PV_DUMP_INIT,
+	KVM_PV_DUMP_CONFIG_STOR_STATE,
+	KVM_PV_DUMP_COMPLETE,
+};
+
+struct kvm_s390_pv_dmp {
+	__u64 subcmd;
+	__u64 buff_addr;
+	__u64 buff_len;
+	__u64 gaddr;		/* For dump storage state */
+	__u64 reserved[4];
+};
+
  enum pv_cmd_info_id {
  	KVM_PV_INFO_VM,
  	KVM_PV_INFO_DUMP,
@@ -1696,6 +1710,7 @@ enum pv_cmd_id {
  	KVM_PV_PREP_RESET,
  	KVM_PV_UNSHARE_ALL,
  	KVM_PV_INFO,
+	KVM_PV_DUMP,
  };
 
  struct kvm_pv_cmd {  
author maintainer
#17 lore raw

PATCH v5 7/10 kvm: s390: Add CPU dump functionality

In reply to Janosch Frank

The previous patch introduced the per-VM dump functions now let's focus on dumping the VCPU state via the newly introduced KVM_S390_PV_CPU_COMMAND ioctl which mirrors the VM UV ioctl and can be extended with new commands later.

---
 arch/s390/kvm/kvm-s390.c | 73 ++++++++++++++++++++++++++++++++++++++++
 arch/s390/kvm/kvm-s390.h |  1 +
 arch/s390/kvm/pv.c       | 16 +++++++++
 include/uapi/linux/kvm.h |  4 +++
 4 files changed, 94 insertions(+)
arch/s390/kvm/kvm-s390.c +73 -0
index 6bf9dd85d50f..c0c848c84552 100644
@@ -5129,6 +5129,52 @@ long kvm_arch_vcpu_async_ioctl(struct file *filp,
 	return -ENOIOCTLCMD;
 }
 
+static int kvm_s390_handle_pv_vcpu_dump(struct kvm_vcpu *vcpu,
+					struct kvm_pv_cmd *cmd)
+{
+	struct kvm_s390_pv_dmp dmp;
+	void *data;
+	int ret;
+
+	/* Dump initialization is a prerequisite */
+	if (!vcpu->kvm->arch.pv.dumping)
+		return -EINVAL;
+
+	if (copy_from_user(&dmp, (__u8 __user *)cmd->data, sizeof(dmp)))
+		return -EFAULT;
+
+	/* We only handle this subcmd right now */
+	if (dmp.subcmd != KVM_PV_DUMP_CPU)
+		return -EINVAL;
+
+	/* CPU dump length is the same as create cpu storage donation. */
+	if (dmp.buff_len != uv_info.guest_cpu_stor_len)
+		return -EINVAL;
+
+	data = vzalloc(uv_info.guest_cpu_stor_len);
+	if (!data)
+		return -ENOMEM;
+
+	ret = kvm_s390_pv_dump_cpu(vcpu, data, &cmd->rc, &cmd->rrc);
+
+	VCPU_EVENT(vcpu, 3, "PROTVIRT DUMP CPU %d rc %x rrc %x",
+		   vcpu->vcpu_id, cmd->rc, cmd->rrc);
+
+	if (ret) {
+		vfree(data);
+		return -EINVAL;
+	}
+
+	/* On success copy over the dump data */
+	if (copy_to_user((__u8 __user *)dmp.buff_addr, data, uv_info.guest_cpu_stor_len)) {
+		vfree(data);
+		return -EFAULT;
+	}
+
+	vfree(data);
+	return 0;
+}
+
 long kvm_arch_vcpu_ioctl(struct file *filp,
 			 unsigned int ioctl, unsigned long arg)
 {
@@ -5293,6 +5339,33 @@ long kvm_arch_vcpu_ioctl(struct file *filp,
 					   irq_state.len);
 		break;
 	}
+	case KVM_S390_PV_CPU_COMMAND: {
+		struct kvm_pv_cmd cmd;
+
+		r = -EINVAL;
+		if (!is_prot_virt_host())
+			break;
+
+		r = -EFAULT;
+		if (copy_from_user(&cmd, argp, sizeof(cmd)))
+			break;
+
+		r = -EINVAL;
+		if (cmd.flags)
+			break;
+
+		/* We only handle this cmd right now */
+		if (cmd.cmd != KVM_PV_DUMP)
+			break;
+
+		r = kvm_s390_handle_pv_vcpu_dump(vcpu, &cmd);
+
+		/* Always copy over UV rc / rrc data */
+		if (copy_to_user((__u8 __user *)argp, &cmd.rc,
+				 sizeof(cmd.rc) + sizeof(cmd.rrc)))
+			r = -EFAULT;
+		break;
+	}
 	default:
 		r = -ENOTTY;
 	}
arch/s390/kvm/kvm-s390.h +1 -0
index 2868dd0bba25..a39815184350 100644
@@ -250,6 +250,7 @@ int kvm_s390_pv_set_sec_parms(struct kvm *kvm, void *hdr, u64 length, u16 *rc,
 int kvm_s390_pv_unpack(struct kvm *kvm, unsigned long addr, unsigned long size,
 		       unsigned long tweak, u16 *rc, u16 *rrc);
 int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state);
+int kvm_s390_pv_dump_cpu(struct kvm_vcpu *vcpu, void *buff, u16 *rc, u16 *rrc);
 int kvm_s390_pv_dump_stor_state(struct kvm *kvm, void __user *buff_user,
 				u64 *gaddr, u64 buff_user_len, u16 *rc, u16 *rrc);
 
arch/s390/kvm/pv.c +16 -0
index fd261667d2c2..8c72330efc0e 100644
@@ -300,6 +300,22 @@ int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state)
 	return 0;
 }
 
+int kvm_s390_pv_dump_cpu(struct kvm_vcpu *vcpu, void *buff, u16 *rc, u16 *rrc)
+{
+	struct uv_cb_dump_cpu uvcb = {
+		.header.cmd = UVC_CMD_DUMP_CPU,
+		.header.len = sizeof(uvcb),
+		.cpu_handle = vcpu->arch.pv.handle,
+		.dump_area_origin = (u64)buff,
+	};
+	int cc;
+
+	cc = uv_call_sched(0, (u64)&uvcb);
+	*rc = uvcb.header.rc;
+	*rrc = uvcb.header.rrc;
+	return cc;
+}
+
 /* Size of the cache for the storage state dump data. 1MB for now */
 #define DUMP_BUFF_LEN HPAGE_SIZE
 
include/uapi/linux/kvm.h +4 -0
index 1c60c2d314ba..0a8b57654ea7 100644
@@ -1657,6 +1657,7 @@ enum pv_cmd_dmp_id {
 	KVM_PV_DUMP_INIT,
 	KVM_PV_DUMP_CONFIG_STOR_STATE,
 	KVM_PV_DUMP_COMPLETE,
+	KVM_PV_DUMP_CPU,
 };
 
 struct kvm_s390_pv_dmp {
@@ -2118,4 +2119,7 @@ struct kvm_stats_desc {
 /* Available with KVM_CAP_XSAVE2 */
 #define KVM_GET_XSAVE2		  _IOR(KVMIO,  0xcf, struct kvm_xsave)
 
+/* Available with KVM_CAP_S390_PROTECTED_DUMP */
+#define KVM_S390_PV_CPU_COMMAND	_IOWR(KVMIO, 0xd0, struct kvm_pv_cmd)
+
 #endif /* __LINUX_KVM_H */
Signature · 1 line
2.34.1
maintainer
#18 lore raw

Re: [PATCH v5 07/10] kvm: s390: Add CPU dump functionality

Quoted 34 lines · On Mon, 16 May 2022 09:08:14 +0000 Janosch Frank <[email protected]> wrote:

The previous patch introduced the per-VM dump functions now let's focus on dumping the VCPU state via the newly introduced KVM_S390_PV_CPU_COMMAND ioctl which mirrors the VM UV ioctl and can be extended with new commands later.

---
 arch/s390/kvm/kvm-s390.c | 73 ++++++++++++++++++++++++++++++++++++++++
 arch/s390/kvm/kvm-s390.h |  1 +
 arch/s390/kvm/pv.c       | 16 +++++++++
 include/uapi/linux/kvm.h |  4 +++
 4 files changed, 94 insertions(+)
arch/s390/kvm/kvm-s390.c +13 -0
index 6bf9dd85d50f..c0c848c84552 100644
@@ -5129,6 +5129,52 @@ long kvm_arch_vcpu_async_ioctl(struct file *filp,
 	return -ENOIOCTLCMD;
 }
 
+static int kvm_s390_handle_pv_vcpu_dump(struct kvm_vcpu *vcpu,
+					struct kvm_pv_cmd *cmd)
+{
+	struct kvm_s390_pv_dmp dmp;
+	void *data;
+	int ret;
+
+	/* Dump initialization is a prerequisite */
+	if (!vcpu->kvm->arch.pv.dumping)
+		return -EINVAL;
+
+	if (copy_from_user(&dmp, (__u8 __user *)cmd->data, sizeof(dmp)))
+		return -EFAULT;

arch/s390/kvm/kvm-s390.c long kvm_arch_vcpu_async_ioctl(struct file *filp, · line 5154

Quoted 10 lines
arch/s390/kvm/kvm-s390.c +10 -0
+
+	/* We only handle this subcmd right now */
+	if (dmp.subcmd != KVM_PV_DUMP_CPU)
+		return -EINVAL;
+
+	/* CPU dump length is the same as create cpu storage donation. */
+	if (dmp.buff_len != uv_info.guest_cpu_stor_len)
+		return -EINVAL;
+
+	data = vzalloc(uv_info.guest_cpu_stor_len);

is it really so big that you need to do a virtual allocation? (I thought it was at most a few pages)

can you at least use kvzalloc?

arch/s390/kvm/kvm-s390.c long kvm_arch_vcpu_async_ioctl(struct file *filp, · line 5165

Quoted 11 lines
arch/s390/kvm/kvm-s390.c +11 -0
+	if (!data)
+		return -ENOMEM;
+
+	ret = kvm_s390_pv_dump_cpu(vcpu, data, &cmd->rc, &cmd->rrc);
+
+	VCPU_EVENT(vcpu, 3, "PROTVIRT DUMP CPU %d rc %x rrc %x",
+		   vcpu->vcpu_id, cmd->rc, cmd->rrc);
+
+	if (ret) {
+		vfree(data);
+		return -EINVAL;

if (ret)

	ret = -EINVAL;

arch/s390/kvm/kvm-s390.c long kvm_arch_vcpu_async_ioctl(struct file *filp, · line 5169

Quoted 4 lines
arch/s390/kvm/kvm-s390.c +4 -0
+	}
+
+	/* On success copy over the dump data */
+	if (copy_to_user((__u8 __user *)dmp.buff_addr, data, uv_info.guest_cpu_stor_len)) {

if (!ret && ...)

	ret = -EFAULT;

arch/s390/kvm/kvm-s390.c long kvm_arch_vcpu_async_ioctl(struct file *filp, · line 5175

Quoted 6 lines
arch/s390/kvm/kvm-s390.c +6 -0
+		vfree(data);
+		return -EFAULT;
+	}
+
+	vfree(data);
+	return 0;

return ret;

this way you have only one free and one return, should be more readable

Quoted 98 lines
arch/s390/kvm/kvm-s390.c +29 -0
+}
+
 long kvm_arch_vcpu_ioctl(struct file *filp,
 			 unsigned int ioctl, unsigned long arg)
 {
@@ -5293,6 +5339,33 @@ long kvm_arch_vcpu_ioctl(struct file *filp,
 					   irq_state.len);
 		break;
 	}
+	case KVM_S390_PV_CPU_COMMAND: {
+		struct kvm_pv_cmd cmd;
+
+		r = -EINVAL;
+		if (!is_prot_virt_host())
+			break;
+
+		r = -EFAULT;
+		if (copy_from_user(&cmd, argp, sizeof(cmd)))
+			break;
+
+		r = -EINVAL;
+		if (cmd.flags)
+			break;
+
+		/* We only handle this cmd right now */
+		if (cmd.cmd != KVM_PV_DUMP)
+			break;
+
+		r = kvm_s390_handle_pv_vcpu_dump(vcpu, &cmd);
+
+		/* Always copy over UV rc / rrc data */
+		if (copy_to_user((__u8 __user *)argp, &cmd.rc,
+				 sizeof(cmd.rc) + sizeof(cmd.rrc)))
+			r = -EFAULT;
+		break;
+	}
 	default:
 		r = -ENOTTY;
 	}
arch/s390/kvm/kvm-s390.h +1 -0
index 2868dd0bba25..a39815184350 100644
@@ -250,6 +250,7 @@ int kvm_s390_pv_set_sec_parms(struct kvm *kvm, void *hdr, u64 length, u16 *rc,
 int kvm_s390_pv_unpack(struct kvm *kvm, unsigned long addr, unsigned long size,
 		       unsigned long tweak, u16 *rc, u16 *rrc);
 int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state);
+int kvm_s390_pv_dump_cpu(struct kvm_vcpu *vcpu, void *buff, u16 *rc, u16 *rrc);
 int kvm_s390_pv_dump_stor_state(struct kvm *kvm, void __user *buff_user,
 				u64 *gaddr, u64 buff_user_len, u16 *rc, u16 *rrc);
 
arch/s390/kvm/pv.c +16 -0
index fd261667d2c2..8c72330efc0e 100644
@@ -300,6 +300,22 @@ int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state)
 	return 0;
 }
 
+int kvm_s390_pv_dump_cpu(struct kvm_vcpu *vcpu, void *buff, u16 *rc, u16 *rrc)
+{
+	struct uv_cb_dump_cpu uvcb = {
+		.header.cmd = UVC_CMD_DUMP_CPU,
+		.header.len = sizeof(uvcb),
+		.cpu_handle = vcpu->arch.pv.handle,
+		.dump_area_origin = (u64)buff,
+	};
+	int cc;
+
+	cc = uv_call_sched(0, (u64)&uvcb);
+	*rc = uvcb.header.rc;
+	*rrc = uvcb.header.rrc;
+	return cc;
+}
+
 /* Size of the cache for the storage state dump data. 1MB for now */
 #define DUMP_BUFF_LEN HPAGE_SIZE
 
include/uapi/linux/kvm.h +4 -0
index 1c60c2d314ba..0a8b57654ea7 100644
@@ -1657,6 +1657,7 @@ enum pv_cmd_dmp_id {
 	KVM_PV_DUMP_INIT,
 	KVM_PV_DUMP_CONFIG_STOR_STATE,
 	KVM_PV_DUMP_COMPLETE,
+	KVM_PV_DUMP_CPU,
 };
 
 struct kvm_s390_pv_dmp {
@@ -2118,4 +2119,7 @@ struct kvm_stats_desc {
 /* Available with KVM_CAP_XSAVE2 */
 #define KVM_GET_XSAVE2		  _IOR(KVMIO,  0xcf, struct kvm_xsave)
 
+/* Available with KVM_CAP_S390_PROTECTED_DUMP */
+#define KVM_S390_PV_CPU_COMMAND	_IOWR(KVMIO, 0xd0, struct kvm_pv_cmd)
+
 #endif /* __LINUX_KVM_H */
author maintainer
#19 lore raw

Re: [PATCH v5 07/10] kvm: s390: Add CPU dump functionality

Quoted 42 lines · Claudio Imbrenda
Quoted 39 lines · On Mon, 16 May 2022 09:08:14 +0000 Janosch Frank <[email protected]> wrote:

The previous patch introduced the per-VM dump functions now let's focus on dumping the VCPU state via the newly introduced KVM_S390_PV_CPU_COMMAND ioctl which mirrors the VM UV ioctl and can be extended with new commands later.

---
  arch/s390/kvm/kvm-s390.c | 73 ++++++++++++++++++++++++++++++++++++++++
  arch/s390/kvm/kvm-s390.h |  1 +
  arch/s390/kvm/pv.c       | 16 +++++++++
  include/uapi/linux/kvm.h |  4 +++
  4 files changed, 94 insertions(+)
arch/s390/kvm/kvm-s390.c +18 -0
index 6bf9dd85d50f..c0c848c84552 100644
@@ -5129,6 +5129,52 @@ long kvm_arch_vcpu_async_ioctl(struct file *filp,
  	return -ENOIOCTLCMD;
  }
 
+static int kvm_s390_handle_pv_vcpu_dump(struct kvm_vcpu *vcpu,
+					struct kvm_pv_cmd *cmd)
+{
+	struct kvm_s390_pv_dmp dmp;
+	void *data;
+	int ret;
+
+	/* Dump initialization is a prerequisite */
+	if (!vcpu->kvm->arch.pv.dumping)
+		return -EINVAL;
+
+	if (copy_from_user(&dmp, (__u8 __user *)cmd->data, sizeof(dmp)))
+		return -EFAULT;
+
+	/* We only handle this subcmd right now */
+	if (dmp.subcmd != KVM_PV_DUMP_CPU)
+		return -EINVAL;
+
Quoted 8 lines

arch/s390/kvm/kvm-s390.c long kvm_arch_vcpu_async_ioctl(struct file *filp, · line 5154

Quoted 5 lines
arch/s390/kvm/kvm-s390.c +5 -0
+	/* CPU dump length is the same as create cpu storage donation. */
+	if (dmp.buff_len != uv_info.guest_cpu_stor_len)
+		return -EINVAL;
+
+	data = vzalloc(uv_info.guest_cpu_stor_len);

is it really so big that you need to do a virtual allocation? (I thought it was at most a few pages)

It's at least a page

Quoted 2 lines

can you at least use kvzalloc?

Sure

Quoted 135 lines

arch/s390/kvm/kvm-s390.c long kvm_arch_vcpu_async_ioctl(struct file *filp, · line 5165

Quoted 12 lines
arch/s390/kvm/kvm-s390.c +11 -0
+	if (!data)
+		return -ENOMEM;
+
+	ret = kvm_s390_pv_dump_cpu(vcpu, data, &cmd->rc, &cmd->rrc);
+
+	VCPU_EVENT(vcpu, 3, "PROTVIRT DUMP CPU %d rc %x rrc %x",
+		   vcpu->vcpu_id, cmd->rc, cmd->rrc);
+
+	if (ret) {
+		vfree(data);
+		return -EINVAL;

if (ret)

	ret = -EINVAL;

arch/s390/kvm/kvm-s390.c long kvm_arch_vcpu_async_ioctl(struct file *filp, · line 5169

Quoted 4 lines
arch/s390/kvm/kvm-s390.c +4 -0
+	}
+
+	/* On success copy over the dump data */
+	if (copy_to_user((__u8 __user *)dmp.buff_addr, data, uv_info.guest_cpu_stor_len)) {

if (!ret && ...)

	ret = -EFAULT;

arch/s390/kvm/kvm-s390.c long kvm_arch_vcpu_async_ioctl(struct file *filp, · line 5175

Quoted 6 lines
arch/s390/kvm/kvm-s390.c +6 -0
+		vfree(data);
+		return -EFAULT;
+	}
+
+	vfree(data);
+	return 0;

return ret;

this way you have only one free and one return, should be more readable

Quoted 98 lines
arch/s390/kvm/kvm-s390.c +29 -0
+}
+
  long kvm_arch_vcpu_ioctl(struct file *filp,
  			 unsigned int ioctl, unsigned long arg)
  {
@@ -5293,6 +5339,33 @@ long kvm_arch_vcpu_ioctl(struct file *filp,
  					   irq_state.len);
  		break;
  	}
+	case KVM_S390_PV_CPU_COMMAND: {
+		struct kvm_pv_cmd cmd;
+
+		r = -EINVAL;
+		if (!is_prot_virt_host())
+			break;
+
+		r = -EFAULT;
+		if (copy_from_user(&cmd, argp, sizeof(cmd)))
+			break;
+
+		r = -EINVAL;
+		if (cmd.flags)
+			break;
+
+		/* We only handle this cmd right now */
+		if (cmd.cmd != KVM_PV_DUMP)
+			break;
+
+		r = kvm_s390_handle_pv_vcpu_dump(vcpu, &cmd);
+
+		/* Always copy over UV rc / rrc data */
+		if (copy_to_user((__u8 __user *)argp, &cmd.rc,
+				 sizeof(cmd.rc) + sizeof(cmd.rrc)))
+			r = -EFAULT;
+		break;
+	}
  	default:
  		r = -ENOTTY;
  	}
arch/s390/kvm/kvm-s390.h +1 -0
index 2868dd0bba25..a39815184350 100644
@@ -250,6 +250,7 @@ int kvm_s390_pv_set_sec_parms(struct kvm *kvm, void *hdr, u64 length, u16 *rc,
  int kvm_s390_pv_unpack(struct kvm *kvm, unsigned long addr, unsigned long size,
  		       unsigned long tweak, u16 *rc, u16 *rrc);
  int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state);
+int kvm_s390_pv_dump_cpu(struct kvm_vcpu *vcpu, void *buff, u16 *rc, u16 *rrc);
  int kvm_s390_pv_dump_stor_state(struct kvm *kvm, void __user *buff_user,
  				u64 *gaddr, u64 buff_user_len, u16 *rc, u16 *rrc);
 
arch/s390/kvm/pv.c +16 -0
index fd261667d2c2..8c72330efc0e 100644
@@ -300,6 +300,22 @@ int kvm_s390_pv_set_cpu_state(struct kvm_vcpu *vcpu, u8 state)
  	return 0;
  }
 
+int kvm_s390_pv_dump_cpu(struct kvm_vcpu *vcpu, void *buff, u16 *rc, u16 *rrc)
+{
+	struct uv_cb_dump_cpu uvcb = {
+		.header.cmd = UVC_CMD_DUMP_CPU,
+		.header.len = sizeof(uvcb),
+		.cpu_handle = vcpu->arch.pv.handle,
+		.dump_area_origin = (u64)buff,
+	};
+	int cc;
+
+	cc = uv_call_sched(0, (u64)&uvcb);
+	*rc = uvcb.header.rc;
+	*rrc = uvcb.header.rrc;
+	return cc;
+}
+
  /* Size of the cache for the storage state dump data. 1MB for now */
  #define DUMP_BUFF_LEN HPAGE_SIZE
 
include/uapi/linux/kvm.h +4 -0
index 1c60c2d314ba..0a8b57654ea7 100644
@@ -1657,6 +1657,7 @@ enum pv_cmd_dmp_id {
  	KVM_PV_DUMP_INIT,
  	KVM_PV_DUMP_CONFIG_STOR_STATE,
  	KVM_PV_DUMP_COMPLETE,
+	KVM_PV_DUMP_CPU,
  };
 
  struct kvm_s390_pv_dmp {
@@ -2118,4 +2119,7 @@ struct kvm_stats_desc {
  /* Available with KVM_CAP_XSAVE2 */
  #define KVM_GET_XSAVE2		  _IOR(KVMIO,  0xcf, struct kvm_xsave)
 
+/* Available with KVM_CAP_S390_PROTECTED_DUMP */
+#define KVM_S390_PV_CPU_COMMAND	_IOWR(KVMIO, 0xd0, struct kvm_pv_cmd)
+
  #endif /* __LINUX_KVM_H */
author maintainer
#20 lore raw

PATCH v5 8/10 kvm: s390: Add KVM_CAP_S390_PROTECTED_DUMP

In reply to Janosch Frank

The capability indicates dump support for protected VMs.

---
 arch/s390/kvm/kvm-s390.c | 20 ++++++++++++++++++++
 include/uapi/linux/kvm.h |  1 +
 2 files changed, 21 insertions(+)
arch/s390/kvm/kvm-s390.c +20 -0
index c0c848c84552..1d65235ed3d3 100644
@@ -606,6 +606,26 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 	case KVM_CAP_S390_PROTECTED:
 		r = is_prot_virt_host();
 		break;
+	case KVM_CAP_S390_PROTECTED_DUMP: {
+		u64 pv_cmds_dump[] = {
+			BIT_UVC_CMD_DUMP_INIT,
+			BIT_UVC_CMD_DUMP_CONFIG_STOR_STATE,
+			BIT_UVC_CMD_DUMP_CPU,
+			BIT_UVC_CMD_DUMP_COMPLETE,
+		};
+		int i;
+
+		if (!is_prot_virt_host())
+			return 0;
+
+		r = 1;
+		for (i = 0; i < ARRAY_SIZE(pv_cmds_dump); i++) {
+			if (!test_bit_inv(pv_cmds_dump[i],
+					  (unsigned long *)&uv_info.inst_calls_list))
+				return 0;
+		}
+		break;
+	}
 	default:
 		r = 0;
 	}
include/uapi/linux/kvm.h +1 -0
index 0a8b57654ea7..ba8f2985a8c0 100644
@@ -1152,6 +1152,7 @@ struct kvm_ppc_resize_hpt {
 #define KVM_CAP_DISABLE_QUIRKS2 213
 /* #define KVM_CAP_VM_TSC_CONTROL 214 */
 #define KVM_CAP_SYSTEM_EVENT_DATA 215
+#define KVM_CAP_S390_PROTECTED_DUMP 216
 
 #ifdef KVM_CAP_IRQ_ROUTING
 
Signature · 1 line
2.34.1
maintainer
#21 lore raw

Re: [PATCH v5 08/10] kvm: s390: Add KVM_CAP_S390_PROTECTED_DUMP

Quoted 17 lines · On Mon, 16 May 2022 09:08:15 +0000 Janosch Frank <[email protected]> wrote:

The capability indicates dump support for protected VMs.

---
 arch/s390/kvm/kvm-s390.c | 20 ++++++++++++++++++++
 include/uapi/linux/kvm.h |  1 +
 2 files changed, 21 insertions(+)
arch/s390/kvm/kvm-s390.c +1 -0
index c0c848c84552..1d65235ed3d3 100644
@@ -606,6 +606,26 @@ int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext)
 	case KVM_CAP_S390_PROTECTED:
 		r = is_prot_virt_host();
 		break;
+	case KVM_CAP_S390_PROTECTED_DUMP: {

arch/s390/kvm/kvm-s390.c int kvm_vm_ioctl_check_extension(struct kvm *kvm, long ext) · line 619

Quoted 10 lines
arch/s390/kvm/kvm-s390.c +10 -0
+		u64 pv_cmds_dump[] = {
+			BIT_UVC_CMD_DUMP_INIT,
+			BIT_UVC_CMD_DUMP_CONFIG_STOR_STATE,
+			BIT_UVC_CMD_DUMP_CPU,
+			BIT_UVC_CMD_DUMP_COMPLETE,
+		};
+		int i;
+
+		if (!is_prot_virt_host())
+			return 0;

we don't return usually in this function, can you use break instad? or maybe even something simpler like this:

r = is_prot_virt_host();

for (i = 0; r && i < ARRAY_SIZE(...); i++)

	r = r && test_bit_inv(...);
Quoted 24 lines
arch/s390/kvm/kvm-s390.c +9 -0
+
+		r = 1;
+		for (i = 0; i < ARRAY_SIZE(pv_cmds_dump); i++) {
+			if (!test_bit_inv(pv_cmds_dump[i],
+					  (unsigned long *)&uv_info.inst_calls_list))
+				return 0;
+		}
+		break;
+	}
 	default:
 		r = 0;
 	}
include/uapi/linux/kvm.h +1 -0
index 0a8b57654ea7..ba8f2985a8c0 100644
@@ -1152,6 +1152,7 @@ struct kvm_ppc_resize_hpt {
 #define KVM_CAP_DISABLE_QUIRKS2 213
 /* #define KVM_CAP_VM_TSC_CONTROL 214 */
 #define KVM_CAP_SYSTEM_EVENT_DATA 215
+#define KVM_CAP_S390_PROTECTED_DUMP 216
 
 #ifdef KVM_CAP_IRQ_ROUTING
 
author maintainer
#22 lore raw

PATCH v5 9/10 Documentation: virt: Protected virtual machine dumps

In reply to Janosch Frank

Let's add a documentation file which describes the dump process. Since we only copy the UV dump data from the UV to userspace we'll not go into detail here and let the party which processes the data describe its structure.

---
 Documentation/virt/kvm/s390/index.rst        |  1 +
 Documentation/virt/kvm/s390/s390-pv-dump.rst | 64 ++++++++++++++++++++
 2 files changed, 65 insertions(+)
 create mode 100644 Documentation/virt/kvm/s390/s390-pv-dump.rst
Documentation/virt/kvm/s390/index.rst +1 -0
index 605f488f0cc5..44ec9ab14b59 100644
@@ -10,3 +10,4 @@ KVM for s390 systems
    s390-diag
    s390-pv
    s390-pv-boot
+   s390-pv-dump
Documentation/virt/kvm/s390/s390-pv-dump.rst new file +64 -0

64 lines not on this page · raw message

Signature · 1 line
2.34.1
maintainer Acked-by on v5 9/10
#23 lore raw

Re: [PATCH v5 09/10] Documentation: virt: Protected virtual machine dumps

Quoted 6 lines · On Mon, 16 May 2022 09:08:16 +0000 Janosch Frank <[email protected]> wrote:

Let's add a documentation file which describes the dump process. Since we only copy the UV dump data from the UV to userspace we'll not go into detail here and let the party which processes the data describe its structure.

Quoted 85 lines
---
 Documentation/virt/kvm/s390/index.rst        |  1 +
 Documentation/virt/kvm/s390/s390-pv-dump.rst | 64 ++++++++++++++++++++
 2 files changed, 65 insertions(+)
 create mode 100644 Documentation/virt/kvm/s390/s390-pv-dump.rst
Documentation/virt/kvm/s390/index.rst +1 -0
index 605f488f0cc5..44ec9ab14b59 100644
@@ -10,3 +10,4 @@ KVM for s390 systems
    s390-diag
    s390-pv
    s390-pv-boot
+   s390-pv-dump
Documentation/virt/kvm/s390/s390-pv-dump.rst new file +64 -0
new file mode 100644
index 000000000000..e542f06048f3
@@ -0,0 +1,64 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+===========================================
+s390 (IBM Z) Protected Virtualization dumps
+===========================================
+
+Summary
+-------
+
+Dumping a VM is an essential tool for debugging problems inside
+it. This is especially true when a protected VM runs into trouble as
+there's no way to access its memory and registers from the outside
+while it's running.
+
+However when dumping a protected VM we need to maintain its
+confidentiality until the dump is in the hands of the VM owner who
+should be the only one capable of analysing it.
+
+The confidentiality of the VM dump is ensured by the Ultravisor who
+provides an interface to KVM over which encrypted CPU and memory data
+can be requested. The encryption is based on the Customer
+Communication Key which is the key that's used to encrypt VM data in a
+way that the customer is able to decrypt.
+
+
+Dump process
+------------
+
+A dump is done in 3 steps:
+
+**Initiation**
+
+This step initializes the dump process, generates cryptographic seeds
+and extracts dump keys with which the VM dump data will be encrypted.
+
+**Data gathering**
+
+Currently there are two types of data that can be gathered from a VM:
+the memory and the vcpu state.
+
+The vcpu state contains all the important registers, general, floating
+point, vector, control and tod/timers of a vcpu. The vcpu dump can
+contain incomplete data if a vcpu is dumped while an instruction is
+emulated with help of the hypervisor. This is indicated by a flag bit
+in the dump data. For the same reason it is very important to not only
+write out the encrypted vcpu state, but also the unencrypted state
+from the hypervisor.
+
+The memory state is further divided into the encrypted memory and its
+metadata comprised of the encryption tweaks and status flags. The
+encrypted memory can simply be read once it has been exported. The
+time of the export does not matter as no re-encryption is
+needed. Memory that has been swapped out and hence was exported can be
+read from the swap and written to the dump target without need for any
+special actions.
+
+The tweaks / status flags for the exported pages need to be requested
+from the Ultravisor.
+
+**Finalization**
+
+The finalization step will provide the data needed to be able to
+decrypt the vcpu and memory data and end the dump process. When this
+step completes successfully a new dump initiation can be started.
author maintainer
#24 lore raw

PATCH v5 10/10 Documentation/virt/kvm/api.rst: Add protvirt dump/info api descriptions

In reply to Janosch Frank

Time to add the dump API changes to the api documentation file. Also some minor cleanup.

---
 Documentation/virt/kvm/api.rst | 153 ++++++++++++++++++++++++++++++++-
 1 file changed, 151 insertions(+), 2 deletions(-)
Documentation/virt/kvm/api.rst +151 -2
index 4a900cdbc62e..c7c964887f5f 100644
@@ -5061,7 +5061,7 @@ into ESA mode. This reset is a superset of the initial reset.
 	__u32 reserved[3];
   };
 
-cmd values:
+**cmd values:**
 
 KVM_PV_ENABLE
   Allocate memory and register the VM with the Ultravisor, thereby
@@ -5077,7 +5077,6 @@ KVM_PV_ENABLE
   =====      =============================
 
 KVM_PV_DISABLE
-
   Deregister the VM from the Ultravisor and reclaim the memory that
   had been donated to the Ultravisor, making it usable by the kernel
   again.  All registered VCPUs are converted back to non-protected
@@ -5094,6 +5093,115 @@ KVM_PV_VM_VERIFY
   Verify the integrity of the unpacked image. Only if this succeeds,
   KVM is allowed to start protected VCPUs.
 
+KVM_PV_INFO
+  :Capability: KVM_CAP_S390_PROTECTED_DUMP
+
+  Presents an API that provides Ultravisor related data to userspace
+  via subcommands. len_max is the size of the user space buffer,
+  len_written is KVM's indication of how much bytes of that buffer
+  were actually written to. len_written can be used to determine the
+  valid fields if more response fields are added in the future.
+
+  ::
+
+     enum pv_cmd_info_id {
+        KVM_PV_INFO_VM,
+        KVM_PV_INFO_DUMP,
+     };
+
+     struct kvm_s390_pv_info_header {
+        __u32 id;
+        __u32 len_max;
+        __u32 len_written;
+        __u32 reserved;
+     };
+
+     struct kvm_s390_pv_info {
+        struct kvm_s390_pv_info_header header;
+        struct kvm_s390_pv_info_dump dump;
+	struct kvm_s390_pv_info_vm vm;
+     };
+
+**subcommands:**
+
+  KVM_PV_INFO_VM
+    This subcommand provides basic Ultravisor information for PV
+    hosts. These values are likely also exported as files in the sysfs
+    firmware UV query interface but they are more easily available to
+    programs in this API.
+
+    The installed calls and feature_indication members provide the
+    installed UV calls and the UV's other feature indications.
+
+    The max_* members provide information about the maximum number of PV
+    vcpus, PV guests and PV guest memory size.
+
+    ::
+
+      struct kvm_s390_pv_info_vm {
+        __u64 inst_calls_list[4];
+        __u64 max_cpus;
+        __u64 max_guests;
+        __u64 max_guest_addr;
+        __u64 feature_indication;
+      };
+
+
+  KVM_PV_INFO_DUMP
+    This subcommand provides information related to dumping PV guests.
+
+    ::
+
+      struct kvm_s390_pv_info_dump {
+        __u64 dump_cpu_buffer_len;
+        __u64 dump_config_mem_buffer_per_1m;
+        __u64 dump_config_finalize_len;
+      };
+
+KVM_PV_DUMP
+  :Capability: KVM_CAP_S390_PROTECTED_DUMP
+
+  Presents an API that provides calls which facilitate dumping a
+  protected VM.
+
+  ::
+
+    struct kvm_s390_pv_dmp {
+      __u64 subcmd;
+      __u64 buff_addr;
+      __u64 buff_len;
+      __u64 gaddr;		/* For dump storage state */
+    };
+
+  **subcommands:**
+
+  KVM_PV_DUMP_INIT
+    Initializes the dump process of a protected VM. If this call does
+    not succeed all other subcommands will fail with -EINVAL. This
+    subcommand will return -EINVAL if a dump process has not yet been
+    completed.
+
+    Not all PV vms can be dumped, the owner needs to set `dump
+    allowed` PCF bit 34 in the SE header to allow dumping.
+
+  KVM_PV_DUMP_CONFIG_STOR_STATE
+    Stores `buff_len` bytes of tweak component values starting with
+    the 1MB block specified by the absolute guest address
+    (`gaddr`). `buff_len` needs to be `conf_dump_storage_state_len`
+    aligned and at least >= the `conf_dump_storage_state_len` value
+    provided by the dump uv_info data.
+
+  KVM_PV_DUMP_COMPLETE
+    If the subcommand succeeds it completes the dump process and lets
+    KVM_PV_DUMP_INIT be called again.
+
+    On success `conf_dump_finalize_len` bytes of completion data will be
+    stored to the `buff_addr`. The completion data contains a key
+    derivation seed, IV, tweak nonce and encryption keys as well as an
+    authentication tag all of which are needed to decrypt the dump at a
+    later time.
+
+
 4.126 KVM_X86_SET_MSR_FILTER
 ----------------------------
 
@@ -5646,6 +5754,32 @@ The offsets of the state save areas in struct kvm_xsave follow the contents
 of CPUID leaf 0xD on the host.
 
 
+4.135 KVM_S390_PV_CPU_COMMAND
+-----------------------------
+
+:Capability: KVM_CAP_S390_PROTECTED_DUMP
+:Architectures: s390
+:Type: vcpu ioctl
+:Parameters: none
+:Returns: 0 on success, < 0 on error
+
+This ioctl closely mirrors `KVM_S390_PV_COMMAND` but handles requests
+for vcpus. It re-uses the kvm_s390_pv_dmp struct and hence also shares
+the command ids.
+
+**command:**
+
+KVM_PV_DUMP
+  Presents an API that provides calls which facilitate dumping a vcpu
+  of a protected VM.
+
+**subcommand:**
+
+KVM_PV_DUMP_CPU
+  Provides encrypted dump data like register values.
+  The length of the returned data is provided by uv_info.guest_cpu_stor_len.
+
+
 5. The kvm_run structure
 ========================
 
@@ -7734,6 +7868,21 @@ At this time, KVM_PMU_CAP_DISABLE is the only capability.  Setting
 this capability will disable PMU virtualization for that VM.  Usermode
 should adjust CPUID leaf 0xA to reflect that the PMU is disabled.
 
+
+8.36 KVM_CAP_S390_PROTECTED_DUMP
+--------------------------------
+
+:Capability: KVM_CAP_S390_PROTECTED_DUMP
+:Architectures: s390
+:Type: vm
+
+This capability indicates that KVM and the Ultravisor support dumping
+PV guests. The `KVM_PV_DUMP` command is available for the
+`KVM_S390_PV_COMMAND` ioctl and the `KVM_PV_INFO` command provides
+dump related UV data. Also the vcpu ioctl `KVM_S390_PV_CPU_COMMAND` is
+available and supports the `KVM_PV_DUMP_CPU` subcommand.
+
+
 9. Known KVM API problems
 =========================
 
Signature · 1 line
2.34.1
maintainer
#25 lore raw

Re: [PATCH v5 10/10] Documentation/virt/kvm/api.rst: Add protvirt dump/info api descriptions

Quoted 120 lines · On Mon, 16 May 2022 09:08:17 +0000 Janosch Frank <[email protected]> wrote:

Time to add the dump API changes to the api documentation file. Also some minor cleanup.

---
 Documentation/virt/kvm/api.rst | 153 ++++++++++++++++++++++++++++++++-
 1 file changed, 151 insertions(+), 2 deletions(-)
Documentation/virt/kvm/api.rst +88 -2
index 4a900cdbc62e..c7c964887f5f 100644
@@ -5061,7 +5061,7 @@ into ESA mode. This reset is a superset of the initial reset.
 	__u32 reserved[3];
   };
 
-cmd values:
+**cmd values:**
 
 KVM_PV_ENABLE
   Allocate memory and register the VM with the Ultravisor, thereby
@@ -5077,7 +5077,6 @@ KVM_PV_ENABLE
   =====      =============================
 
 KVM_PV_DISABLE
-
   Deregister the VM from the Ultravisor and reclaim the memory that
   had been donated to the Ultravisor, making it usable by the kernel
   again.  All registered VCPUs are converted back to non-protected
@@ -5094,6 +5093,115 @@ KVM_PV_VM_VERIFY
   Verify the integrity of the unpacked image. Only if this succeeds,
   KVM is allowed to start protected VCPUs.
 
+KVM_PV_INFO
+  :Capability: KVM_CAP_S390_PROTECTED_DUMP
+
+  Presents an API that provides Ultravisor related data to userspace
+  via subcommands. len_max is the size of the user space buffer,
+  len_written is KVM's indication of how much bytes of that buffer
+  were actually written to. len_written can be used to determine the
+  valid fields if more response fields are added in the future.
+
+  ::
+
+     enum pv_cmd_info_id {
+        KVM_PV_INFO_VM,
+        KVM_PV_INFO_DUMP,
+     };
+
+     struct kvm_s390_pv_info_header {
+        __u32 id;
+        __u32 len_max;
+        __u32 len_written;
+        __u32 reserved;
+     };
+
+     struct kvm_s390_pv_info {
+        struct kvm_s390_pv_info_header header;
+        struct kvm_s390_pv_info_dump dump;
+	struct kvm_s390_pv_info_vm vm;
+     };
+
+**subcommands:**
+
+  KVM_PV_INFO_VM
+    This subcommand provides basic Ultravisor information for PV
+    hosts. These values are likely also exported as files in the sysfs
+    firmware UV query interface but they are more easily available to
+    programs in this API.
+
+    The installed calls and feature_indication members provide the
+    installed UV calls and the UV's other feature indications.
+
+    The max_* members provide information about the maximum number of PV
+    vcpus, PV guests and PV guest memory size.
+
+    ::
+
+      struct kvm_s390_pv_info_vm {
+        __u64 inst_calls_list[4];
+        __u64 max_cpus;
+        __u64 max_guests;
+        __u64 max_guest_addr;
+        __u64 feature_indication;
+      };
+
+
+  KVM_PV_INFO_DUMP
+    This subcommand provides information related to dumping PV guests.
+
+    ::
+
+      struct kvm_s390_pv_info_dump {
+        __u64 dump_cpu_buffer_len;
+        __u64 dump_config_mem_buffer_per_1m;
+        __u64 dump_config_finalize_len;
+      };
+
+KVM_PV_DUMP
+  :Capability: KVM_CAP_S390_PROTECTED_DUMP
+
+  Presents an API that provides calls which facilitate dumping a
+  protected VM.
+
+  ::
+
+    struct kvm_s390_pv_dmp {
+      __u64 subcmd;
+      __u64 buff_addr;
+      __u64 buff_len;
+      __u64 gaddr;		/* For dump storage state */
+    };
+
+  **subcommands:**
+
+  KVM_PV_DUMP_INIT
+    Initializes the dump process of a protected VM. If this call does
+    not succeed all other subcommands will fail with -EINVAL. This
+    subcommand will return -EINVAL if a dump process has not yet been
+    completed.

Documentation/virt/kvm/api.rst KVM_PV_VM_VERIFY · line 5192

Quoted 10 lines
Documentation/virt/kvm/api.rst +10 -0
+
+    Not all PV vms can be dumped, the owner needs to set `dump
+    allowed` PCF bit 34 in the SE header to allow dumping.
+
+  KVM_PV_DUMP_CONFIG_STOR_STATE
+    Stores `buff_len` bytes of tweak component values starting with
+    the 1MB block specified by the absolute guest address
+    (`gaddr`). `buff_len` needs to be `conf_dump_storage_state_len`
+    aligned and at least >= the `conf_dump_storage_state_len` value
+    provided by the dump uv_info data.

please explain that the output buffer might be written to (even partially) even when the IOCTL fails

Quoted 70 lines
Documentation/virt/kvm/api.rst +53 -0
+
+  KVM_PV_DUMP_COMPLETE
+    If the subcommand succeeds it completes the dump process and lets
+    KVM_PV_DUMP_INIT be called again.
+
+    On success `conf_dump_finalize_len` bytes of completion data will be
+    stored to the `buff_addr`. The completion data contains a key
+    derivation seed, IV, tweak nonce and encryption keys as well as an
+    authentication tag all of which are needed to decrypt the dump at a
+    later time.
+
+
 4.126 KVM_X86_SET_MSR_FILTER
 ----------------------------
 
@@ -5646,6 +5754,32 @@ The offsets of the state save areas in struct kvm_xsave follow the contents
 of CPUID leaf 0xD on the host.
 
 
+4.135 KVM_S390_PV_CPU_COMMAND
+-----------------------------
+
+:Capability: KVM_CAP_S390_PROTECTED_DUMP
+:Architectures: s390
+:Type: vcpu ioctl
+:Parameters: none
+:Returns: 0 on success, < 0 on error
+
+This ioctl closely mirrors `KVM_S390_PV_COMMAND` but handles requests
+for vcpus. It re-uses the kvm_s390_pv_dmp struct and hence also shares
+the command ids.
+
+**command:**
+
+KVM_PV_DUMP
+  Presents an API that provides calls which facilitate dumping a vcpu
+  of a protected VM.
+
+**subcommand:**
+
+KVM_PV_DUMP_CPU
+  Provides encrypted dump data like register values.
+  The length of the returned data is provided by uv_info.guest_cpu_stor_len.
+
+
 5. The kvm_run structure
 ========================
 
@@ -7734,6 +7868,21 @@ At this time, KVM_PMU_CAP_DISABLE is the only capability.  Setting
 this capability will disable PMU virtualization for that VM.  Usermode
 should adjust CPUID leaf 0xA to reflect that the PMU is disabled.
 
+
+8.36 KVM_CAP_S390_PROTECTED_DUMP
+--------------------------------
+
+:Capability: KVM_CAP_S390_PROTECTED_DUMP
+:Architectures: s390
+:Type: vm
+
+This capability indicates that KVM and the Ultravisor support dumping
+PV guests. The `KVM_PV_DUMP` command is available for the
+`KVM_S390_PV_COMMAND` ioctl and the `KVM_PV_INFO` command provides
+dump related UV data. Also the vcpu ioctl `KVM_S390_PV_CPU_COMMAND` is
+available and supports the `KVM_PV_DUMP_CPU` subcommand.
+
+
 9. Known KVM API problems
 =========================