Skip to contentSkip to conversation

target/riscv, KVM: fixes and enhancements

PATCH Superseded Daniel Henrique Barboza 2 participants 31 messages Jun 22, 2023 – Jun 26, 2023 RISC-V

v3 of 9 Superseded Latest is v9 →

lore
Share & tools
lore
#0 lore raw

v3 cover target/riscv, KVM: fixes and enhancements

Hi,

In this new version the highlight is the changes in how we're making riscv_isa_string_ext() work with KVM CPUs. A new patch (patch 15) was added where we eliminate an unneeded priv_check from the function, sparing us from having to assign an env->priv_ver for the 'host' CPU type.

Then, in patch 16, instead of making changes with tcg/kvm extensions lists and so on, we're avoiding the problem altogether with mock/no-op properties. We go through all the extension flags (boolean values only) from riscv_cpu_extensions[] and, if the option wasn't already created by the KVM driver, we assign a no-op set() callback to it. This allows the user to use any property, kvm-only or otherwise, and the KVM guest can safely ignore it.

Last but not the least, patch 11 was changed as requested by Andrew in v2.

Series is rebased on top of Alistair's riscv-to-apply.next.

Patches missing review: 11, 15, 16.

Changes from v2:
- patch 11:
- renamed misa_ext_infos[] to misa_ext_info_arr[]
- mention that misa_ext_cfgs[] also needs update in cpu.h
- misa_ext_info_arr[] moved to cpu.c to be used as extern
- patch 13:
- amended commit msg to clarify that all extensions, MISA Or multi-letter, should be enabled and advertised by KVM
- fixed comment in kvm_riscv_update_cpu_misa_ext()
- patch 15 (new):
- remove priv_ver check from riscv_isa_string_ext()
- patch 16 (former 15):
- reworked: instead of reworking riscv_isa_string_ext(), add mock KVM properties
- v2 link: https://lists.gnu.org/archive/html/qemu-devel/2023-06/msg02790.html

Daniel Henrique Barboza (19):
target/riscv: skip features setup for KVM CPUs hw/riscv/virt.c: skip 'mmu-type' FDT if satp mode not set target/riscv/cpu.c: restrict 'mvendorid' value target/riscv/cpu.c: restrict 'mimpid' value target/riscv/cpu.c: restrict 'marchid' value
target/riscv: use KVM scratch CPUs to init KVM properties target/riscv: read marchid/mimpid in kvm_riscv_init_machine_ids() target/riscv: handle mvendorid/marchid/mimpid for KVM CPUs
linux-headers: Update to v6.4-rc1
target/riscv/kvm.c: init 'misa_ext_mask' with scratch CPU target/riscv/cpu: add misa_ext_info_arr[]
target/riscv: add KVM specific MISA properties target/riscv/kvm.c: update KVM MISA bits target/riscv/kvm.c: add multi-letter extension KVM properties target/riscv/cpu.c: remove priv_ver check from riscv_isa_string_ext() target/riscv/cpu.c: create KVM mock properties target/riscv: update multi-letter extension KVM properties target/riscv/kvm.c: add kvmconfig_get_cfg_addr() helper target/riscv/kvm.c: read/write (cbom|cboz)_blocksize in KVM

 hw/riscv/virt.c                               |  14 +-
 include/standard-headers/linux/const.h        |   2 +-
 include/standard-headers/linux/virtio_blk.h   |  18 +-
 .../standard-headers/linux/virtio_config.h    |   6 +
 include/standard-headers/linux/virtio_net.h   |   1 +
 linux-headers/asm-arm64/kvm.h                 |  33 ++
 linux-headers/asm-riscv/kvm.h                 |  53 +-
 linux-headers/asm-riscv/unistd.h              |   9 +
 linux-headers/asm-s390/unistd_32.h            |   1 +
 linux-headers/asm-s390/unistd_64.h            |   1 +
 linux-headers/asm-x86/kvm.h                   |   3 +
 linux-headers/linux/const.h                   |   2 +-
 linux-headers/linux/kvm.h                     |  12 +-
 linux-headers/linux/psp-sev.h                 |   7 +
 linux-headers/linux/userfaultfd.h             |  17 +-
 target/riscv/cpu.c                            | 282 ++++++++--
 target/riscv/cpu.h                            |  11 +-
 target/riscv/kvm.c                            | 499 +++++++++++++++++-
 target/riscv/kvm_riscv.h                      |   1 +
 19 files changed, 891 insertions(+), 81 deletions(-)
Signature · 1 line
2.41.0
#1 lore raw

PATCH v3 1/19 target/riscv: skip features setup for KVM CPUs

As it is today it's not possible to use '-cpu host' if the RISC-V host has RVH enabled. This is the resulting error:

$ sudo ./qemu/build/qemu-system-riscv64 \
    -machine virt,accel=kvm -m 2G -smp 1 \
    -nographic -snapshot -kernel ./guest_imgs/Image  \
    -initrd ./guest_imgs/rootfs_kvm_riscv64.img \
    -append "earlycon=sbi root=/dev/ram rw" \
    -cpu host
qemu-system-riscv64: H extension requires priv spec 1.12.0

This happens because we're checking for priv spec for all CPUs, and since we're not setting env->priv_ver for the 'host' CPU, it's being default to zero (i.e. PRIV_SPEC_1_10_0).

In reality env->priv_ver does not make sense when running with the KVM 'host' CPU. It's used to gate certain CSRs/extensions during translation to make them unavailable if the hart declares an older spec version. It doesn't have any other use. E.g. OpenSBI version 1.2 retrieves the spec checking if the CSR_MCOUNTEREN, CSR_MCOUNTINHIBIT and CSR_MENVCFG CSRs are available [1].

'priv_ver' is just one example. We're doing a lot of feature validation and setup during riscv_cpu_realize() that it doesn't apply KVM CPUs. Validating the feature set for those CPUs is a KVM problem that should be handled in KVM specific code.

The new riscv_cpu_realize_features() helper contains all validation logic that are not applicable to KVM CPUs. riscv_cpu_realize() verifies if we're dealing with a KVM CPU and, if not, execute the new helper to proceed with the usual realize() logic for all other CPUs.

[1] lib/sbi/sbi_hart.c, hart_detect_features()

---
 target/riscv/cpu.c | 43 +++++++++++++++++++++++++++++++++----------
 1 file changed, 33 insertions(+), 10 deletions(-)
target/riscv/cpu.c +33 -10
index fb8458bf74..e515dde208 100644
@@ -331,6 +331,15 @@ static void set_satp_mode_default_map(RISCVCPU *cpu)
 }
 #endif
 
+static bool riscv_running_kvm(void)
+{
+#ifndef CONFIG_USER_ONLY
+    return kvm_enabled();
+#else
+    return false;
+#endif
+}
+
 static void riscv_any_cpu_init(Object *obj)
 {
     RISCVCPU *cpu = RISCV_CPU(obj);
@@ -1308,20 +1317,12 @@ static void riscv_cpu_validate_misa_priv(CPURISCVState *env, Error **errp)
     }
 }
 
-static void riscv_cpu_realize(DeviceState *dev, Error **errp)
+static void riscv_cpu_realize_features(DeviceState *dev, Error **errp)
 {
-    CPUState *cs = CPU(dev);
     RISCVCPU *cpu = RISCV_CPU(dev);
     CPURISCVState *env = &cpu->env;
-    RISCVCPUClass *mcc = RISCV_CPU_GET_CLASS(dev);
     Error *local_err = NULL;
 
-    cpu_exec_realizefn(cs, &local_err);
-    if (local_err != NULL) {
-        error_propagate(errp, local_err);
-        return;
-    }
-
     riscv_cpu_validate_misa_mxl(cpu, &local_err);
     if (local_err != NULL) {
         error_propagate(errp, local_err);
@@ -1356,7 +1357,7 @@ static void riscv_cpu_realize(DeviceState *dev, Error **errp)
     }
 
 #ifndef CONFIG_USER_ONLY
-    cs->tcg_cflags |= CF_PCREL;
+    CPU(dev)->tcg_cflags |= CF_PCREL;
 
     if (cpu->cfg.ext_sstc) {
         riscv_timer_init(cpu);
@@ -1369,6 +1370,28 @@ static void riscv_cpu_realize(DeviceState *dev, Error **errp)
         }
      }
 #endif
+}
+
+static void riscv_cpu_realize(DeviceState *dev, Error **errp)
+{
+    CPUState *cs = CPU(dev);
+    RISCVCPU *cpu = RISCV_CPU(dev);
+    RISCVCPUClass *mcc = RISCV_CPU_GET_CLASS(dev);
+    Error *local_err = NULL;
+
+    cpu_exec_realizefn(cs, &local_err);
+    if (local_err != NULL) {
+        error_propagate(errp, local_err);
+        return;
+    }
+
+    if (!riscv_running_kvm()) {
+        riscv_cpu_realize_features(dev, &local_err);
+        if (local_err != NULL) {
+            error_propagate(errp, local_err);
+            return;
+        }
+    }
 
     riscv_cpu_finalize_features(cpu, &local_err);
     if (local_err != NULL) {
Signature · 1 line
2.41.0
#2 lore raw

PATCH v3 2/19 hw/riscv/virt.c: skip 'mmu-type' FDT if satp mode not set

In reply to Daniel Henrique Barboza

The absence of a satp mode in riscv_host_cpu_init() is causing the following error:

$ sudo ./qemu/build/qemu-system-riscv64  -machine virt,accel=kvm \
    -m 2G -smp 1  -nographic -snapshot \
    -kernel ./guest_imgs/Image \
    -initrd ./guest_imgs/rootfs_kvm_riscv64.img \
    -append "earlycon=sbi root=/dev/ram rw" \
    -cpu host
**
ERROR:../target/riscv/cpu.c:320:satp_mode_str: code should not be
reached
Bail out! ERROR:../target/riscv/cpu.c:320:satp_mode_str: code should
not be reached
Aborted

The error is triggered from create_fdt_socket_cpus() in hw/riscv/virt.c. It's trying to get satp_mode_str for a NULL cpu->cfg.satp_mode.map.

For this KVM cpu we would need to inherit the satp supported modes from the RISC-V host. At this moment this is not possible because the KVM driver does not support it. And even when it does we can't just let this broken for every other older kernel.

Since mmu-type is not a required node, according to [1], skip the 'mmu-type' FDT node if there's no satp_mode set. We'll revisit this logic when we can get satp information from KVM.

[1] https://github.com/devicetree-org/dt-schema/blob/main/dtschema/schemas/cpu.yaml

---
 hw/riscv/virt.c | 14 +++++++-------
 1 file changed, 7 insertions(+), 7 deletions(-)
hw/riscv/virt.c +7 -7
index 95708d890e..f025a0fcaf 100644
@@ -243,13 +243,13 @@ static void create_fdt_socket_cpus(RISCVVirtState *s, int socket,
             s->soc[socket].hartid_base + cpu);
         qemu_fdt_add_subnode(ms->fdt, cpu_name);
 
-        satp_mode_max = satp_mode_max_from_map(
-            s->soc[socket].harts[cpu].cfg.satp_mode.map);
-        sv_name = g_strdup_printf("riscv,%s",
-                                  satp_mode_str(satp_mode_max, is_32_bit));
-        qemu_fdt_setprop_string(ms->fdt, cpu_name, "mmu-type", sv_name);
-        g_free(sv_name);
-
+        if (cpu_ptr->cfg.satp_mode.supported != 0) {
+            satp_mode_max = satp_mode_max_from_map(cpu_ptr->cfg.satp_mode.map);
+            sv_name = g_strdup_printf("riscv,%s",
+                                      satp_mode_str(satp_mode_max, is_32_bit));
+            qemu_fdt_setprop_string(ms->fdt, cpu_name, "mmu-type", sv_name);
+            g_free(sv_name);
+        }
 
         name = riscv_isa_string(cpu_ptr);
         qemu_fdt_setprop_string(ms->fdt, cpu_name, "riscv,isa", name);
Signature · 1 line
2.41.0
#3 lore raw

PATCH v3 3/19 target/riscv/cpu.c: restrict 'mvendorid' value

In reply to Daniel Henrique Barboza

We're going to change the handling of mvendorid/marchid/mimpid by the KVM driver. Since these are always present in all CPUs let's put the same validation for everyone.

It doesn't make sense to allow 'mvendorid' to be different than it is already set in named (vendor) CPUs. Generic (dynamic) CPUs can have any 'mvendorid' they want.

Change 'mvendorid' to be a class property created via 'object_class_property_add', instead of using the DEFINE_PROP_UINT32() macro. This allow us to define a custom setter for it that will verify, for named CPUs, if mvendorid is different than it is already set by the CPU. This is the error thrown for the 'veyron-v1' CPU if 'mvendorid' is set to an invalid value:

$ qemu-system-riscv64 -M virt -nographic -cpu veyron-v1,mvendorid=2
qemu-system-riscv64: can't apply global veyron-v1-riscv-cpu.mvendorid=2:
    Unable to change veyron-v1-riscv-cpu mvendorid (0x61f)
---
 target/riscv/cpu.c | 38 +++++++++++++++++++++++++++++++++++++-
 1 file changed, 37 insertions(+), 1 deletion(-)
target/riscv/cpu.c +37 -1
index e515dde208..b65958a887 100644
@@ -1738,7 +1738,6 @@ static void riscv_cpu_add_user_properties(Object *obj)
 static Property riscv_cpu_properties[] = {
     DEFINE_PROP_BOOL("debug", RISCVCPU, cfg.debug, true),
 
-    DEFINE_PROP_UINT32("mvendorid", RISCVCPU, cfg.mvendorid, 0),
     DEFINE_PROP_UINT64("marchid", RISCVCPU, cfg.marchid, RISCV_CPU_MARCHID),
     DEFINE_PROP_UINT64("mimpid", RISCVCPU, cfg.mimpid, RISCV_CPU_MIMPID),
 
@@ -1825,6 +1824,40 @@ static const struct TCGCPUOps riscv_tcg_ops = {
 #endif /* !CONFIG_USER_ONLY */
 };
 
+static bool riscv_cpu_is_dynamic(Object *cpu_obj)
+{
+    return object_dynamic_cast(cpu_obj, TYPE_RISCV_DYNAMIC_CPU) != NULL;
+}
+
+static void cpu_set_mvendorid(Object *obj, Visitor *v, const char *name,
+                              void *opaque, Error **errp)
+{
+    bool dynamic_cpu = riscv_cpu_is_dynamic(obj);
+    RISCVCPU *cpu = RISCV_CPU(obj);
+    uint32_t prev_val = cpu->cfg.mvendorid;
+    uint32_t value;
+
+    if (!visit_type_uint32(v, name, &value, errp)) {
+        return;
+    }
+
+    if (!dynamic_cpu && prev_val != value) {
+        error_setg(errp, "Unable to change %s mvendorid (0x%x)",
+                   object_get_typename(obj), prev_val);
+        return;
+    }
+
+    cpu->cfg.mvendorid = value;
+}
+
+static void cpu_get_mvendorid(Object *obj, Visitor *v, const char *name,
+                              void *opaque, Error **errp)
+{
+    bool value = RISCV_CPU(obj)->cfg.mvendorid;
+
+    visit_type_bool(v, name, &value, errp);
+}
+
 static void riscv_cpu_class_init(ObjectClass *c, void *data)
 {
     RISCVCPUClass *mcc = RISCV_CPU_CLASS(c);
@@ -1856,6 +1889,9 @@ static void riscv_cpu_class_init(ObjectClass *c, void *data)
     cc->gdb_get_dynamic_xml = riscv_gdb_get_dynamic_xml;
     cc->tcg_ops = &riscv_tcg_ops;
 
+    object_class_property_add(c, "mvendorid", "uint32", cpu_get_mvendorid,
+                              cpu_set_mvendorid, NULL, NULL);
+
     device_class_set_props(dc, riscv_cpu_properties);
 }
 
Signature · 1 line
2.41.0
#4 lore raw

PATCH v3 4/19 target/riscv/cpu.c: restrict 'mimpid' value

In reply to Daniel Henrique Barboza

Following the same logic used with 'mvendorid' let's also restrict 'mimpid' for named CPUs. Generic CPUs keep setting the value freely.

Note that we're getting rid of the default RISCV_CPU_MARCHID value. The reason is that this is not a good default since it's dynamic, changing with with every QEMU version, regardless of whether the actual implementation of the CPU changed from one QEMU version to the other. Named CPU should set it to a meaningful value instead and generic CPUs can set whatever they want.

This is the error thrown for an invalid 'mimpid' value for the veyron-v1 CPU:

$ ./qemu-system-riscv64 -M virt -nographic -cpu veyron-v1,mimpid=2
qemu-system-riscv64: can't apply global veyron-v1-riscv-cpu.mimpid=2:
    Unable to change veyron-v1-riscv-cpu mimpid (0x111)
---
 target/riscv/cpu.c | 34 ++++++++++++++++++++++++++++++++--
 1 file changed, 32 insertions(+), 2 deletions(-)
target/riscv/cpu.c +32 -2
index b65958a887..33cc28acf1 100644
@@ -42,7 +42,6 @@
 #define RISCV_CPU_MARCHID   ((QEMU_VERSION_MAJOR << 16) | \
                              (QEMU_VERSION_MINOR << 8)  | \
                              (QEMU_VERSION_MICRO))
-#define RISCV_CPU_MIMPID    RISCV_CPU_MARCHID
 
 static const char riscv_single_letter_exts[] = "IEMAFDQCPVH";
 
@@ -1739,7 +1738,6 @@ static Property riscv_cpu_properties[] = {
     DEFINE_PROP_BOOL("debug", RISCVCPU, cfg.debug, true),
 
     DEFINE_PROP_UINT64("marchid", RISCVCPU, cfg.marchid, RISCV_CPU_MARCHID),
-    DEFINE_PROP_UINT64("mimpid", RISCVCPU, cfg.mimpid, RISCV_CPU_MIMPID),
 
 #ifndef CONFIG_USER_ONLY
     DEFINE_PROP_UINT64("resetvec", RISCVCPU, env.resetvec, DEFAULT_RSTVEC),
@@ -1858,6 +1856,35 @@ static void cpu_get_mvendorid(Object *obj, Visitor *v, const char *name,
     visit_type_bool(v, name, &value, errp);
 }
 
+static void cpu_set_mimpid(Object *obj, Visitor *v, const char *name,
+                           void *opaque, Error **errp)
+{
+    bool dynamic_cpu = riscv_cpu_is_dynamic(obj);
+    RISCVCPU *cpu = RISCV_CPU(obj);
+    uint64_t prev_val = cpu->cfg.mimpid;
+    uint64_t value;
+
+    if (!visit_type_uint64(v, name, &value, errp)) {
+        return;
+    }
+
+    if (!dynamic_cpu && prev_val != value) {
+        error_setg(errp, "Unable to change %s mimpid (0x%lx)",
+                   object_get_typename(obj), prev_val);
+        return;
+    }
+
+    cpu->cfg.mimpid = value;
+}
+
+static void cpu_get_mimpid(Object *obj, Visitor *v, const char *name,
+                           void *opaque, Error **errp)
+{
+    bool value = RISCV_CPU(obj)->cfg.mimpid;
+
+    visit_type_bool(v, name, &value, errp);
+}
+
 static void riscv_cpu_class_init(ObjectClass *c, void *data)
 {
     RISCVCPUClass *mcc = RISCV_CPU_CLASS(c);
@@ -1892,6 +1919,9 @@ static void riscv_cpu_class_init(ObjectClass *c, void *data)
     object_class_property_add(c, "mvendorid", "uint32", cpu_get_mvendorid,
                               cpu_set_mvendorid, NULL, NULL);
 
+    object_class_property_add(c, "mimpid", "uint64", cpu_get_mimpid,
+                              cpu_set_mimpid, NULL, NULL);
+
     device_class_set_props(dc, riscv_cpu_properties);
 }
 
Signature · 1 line
2.41.0
#5 lore raw

PATCH v3 5/19 target/riscv/cpu.c: restrict 'marchid' value

In reply to Daniel Henrique Barboza

'marchid' shouldn't be set to a different value as previously set for named CPUs.

For all other CPUs it shouldn't be freely set either - the spec requires that 'marchid' can't have the MSB (most significant bit) set and every other bit set to zero, i.e. 0x80000000 is an invalid 'marchid' value for 32 bit CPUs.

As with 'mimpid', setting a default value based on the current QEMU version is not a good idea because it implies that the CPU implementation changes from one QEMU version to the other. Named CPUs should set 'marchid' to a meaningful value instead, and generic CPUs can set to any valid value.

For the 'veyron-v1' CPU this is the error thrown if 'marchid' is set to a different val:

$ ./build/qemu-system-riscv64 -M virt -nographic -cpu veyron-v1,marchid=0x80000000
qemu-system-riscv64: can't apply global veyron-v1-riscv-cpu.marchid=0x80000000:
    Unable to change veyron-v1-riscv-cpu marchid (0x8000000000010000)

And, for generics CPUs, this is the error when trying to set to an invalid val:

$ ./build/qemu-system-riscv64 -M virt -nographic -cpu rv64,marchid=0x8000000000000000
qemu-system-riscv64: can't apply global rv64-riscv-cpu.marchid=0x8000000000000000:
    Unable to set marchid with MSB (64) bit set and the remaining bits zero
---
 target/riscv/cpu.c | 60 ++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 53 insertions(+), 7 deletions(-)
target/riscv/cpu.c +53 -7
index 33cc28acf1..4768dcdf49 100644
@@ -38,11 +38,6 @@
 #include "tcg/tcg.h"
 
 /* RISC-V CPU definitions */
-
-#define RISCV_CPU_MARCHID   ((QEMU_VERSION_MAJOR << 16) | \
-                             (QEMU_VERSION_MINOR << 8)  | \
-                             (QEMU_VERSION_MICRO))
-
 static const char riscv_single_letter_exts[] = "IEMAFDQCPVH";
 
 struct isa_ext_data {
@@ -1737,8 +1732,6 @@ static void riscv_cpu_add_user_properties(Object *obj)
 static Property riscv_cpu_properties[] = {
     DEFINE_PROP_BOOL("debug", RISCVCPU, cfg.debug, true),
 
-    DEFINE_PROP_UINT64("marchid", RISCVCPU, cfg.marchid, RISCV_CPU_MARCHID),
-
 #ifndef CONFIG_USER_ONLY
     DEFINE_PROP_UINT64("resetvec", RISCVCPU, env.resetvec, DEFAULT_RSTVEC),
 #endif
@@ -1885,6 +1878,56 @@ static void cpu_get_mimpid(Object *obj, Visitor *v, const char *name,
     visit_type_bool(v, name, &value, errp);
 }
 
+static void cpu_set_marchid(Object *obj, Visitor *v, const char *name,
+                            void *opaque, Error **errp)
+{
+    bool dynamic_cpu = riscv_cpu_is_dynamic(obj);
+    RISCVCPU *cpu = RISCV_CPU(obj);
+    uint64_t prev_val = cpu->cfg.marchid;
+    uint64_t value, invalid_val;
+    uint32_t mxlen = 0;
+
+    if (!visit_type_uint64(v, name, &value, errp)) {
+        return;
+    }
+
+    if (!dynamic_cpu && prev_val != value) {
+        error_setg(errp, "Unable to change %s marchid (0x%lx)",
+                   object_get_typename(obj), prev_val);
+        return;
+    }
+
+    switch (riscv_cpu_mxl(&cpu->env)) {
+    case MXL_RV32:
+        mxlen = 32;
+        break;
+    case MXL_RV64:
+    case MXL_RV128:
+        mxlen = 64;
+        break;
+    default:
+        g_assert_not_reached();
+    }
+
+    invalid_val = 1LL << (mxlen - 1);
+
+    if (value == invalid_val) {
+        error_setg(errp, "Unable to set marchid with MSB (%u) bit set "
+                         "and the remaining bits zero", mxlen);
+        return;
+    }
+
+    cpu->cfg.marchid = value;
+}
+
+static void cpu_get_marchid(Object *obj, Visitor *v, const char *name,
+                           void *opaque, Error **errp)
+{
+    bool value = RISCV_CPU(obj)->cfg.marchid;
+
+    visit_type_bool(v, name, &value, errp);
+}
+
 static void riscv_cpu_class_init(ObjectClass *c, void *data)
 {
     RISCVCPUClass *mcc = RISCV_CPU_CLASS(c);
@@ -1922,6 +1965,9 @@ static void riscv_cpu_class_init(ObjectClass *c, void *data)
     object_class_property_add(c, "mimpid", "uint64", cpu_get_mimpid,
                               cpu_set_mimpid, NULL, NULL);
 
+    object_class_property_add(c, "marchid", "uint64", cpu_get_marchid,
+                              cpu_set_marchid, NULL, NULL);
+
     device_class_set_props(dc, riscv_cpu_properties);
 }
 
Signature · 1 line
2.41.0
#6 lore raw

PATCH v3 6/19 target/riscv: use KVM scratch CPUs to init KVM properties

In reply to Daniel Henrique Barboza

Certain validations, such as the validations done for the machine IDs (mvendorid/marchid/mimpid), are done before starting the CPU. Non-dynamic (named) CPUs tries to match user input with a preset default. As it is today we can't prefetch a KVM default for these cases because we're only able to read/write KVM regs after the vcpu is spinning.

Our target/arm friends use a concept called "scratch CPU", which consists of creating a vcpu for doing queries and validations and so on, which is discarded shortly after use [1]. This is a suitable solution for what we need so let's implement it in target/riscv as well.

kvm_riscv_init_machine_ids() will be used to do any pre-launch setup for KVM CPUs, via riscv_cpu_add_user_properties(). The function will create a KVM scratch CPU, fetch KVM regs that work as default values for user properties, and then discard the scratch CPU afterwards.

We're starting by initializing 'mvendorid'. This concept will be used to init other KVM specific properties in the next patches as well.

[1] target/arm/kvm.c, kvm_arm_create_scratch_host_vcpu()

---
 target/riscv/cpu.c       |  4 ++
 target/riscv/kvm.c       | 85 ++++++++++++++++++++++++++++++++++++++++
 target/riscv/kvm_riscv.h |  1 +
 3 files changed, 90 insertions(+)
target/riscv/cpu.c +4 -0
index 4768dcdf49..641bec3573 100644
@@ -1718,6 +1718,10 @@ static void riscv_cpu_add_user_properties(Object *obj)
     Property *prop;
     DeviceState *dev = DEVICE(obj);
 
+    if (riscv_running_kvm()) {
+        kvm_riscv_init_user_properties(obj);
+    }
+
     riscv_cpu_add_misa_properties(obj);
 
     for (prop = riscv_cpu_extensions; prop && prop->name; prop++) {
target/riscv/kvm.c +85 -0

91 lines not on this page · raw message

target/riscv/kvm_riscv.h +1 -0
index ed281bdce0..e3ba935808 100644
@@ -19,6 +19,7 @@
 #ifndef QEMU_KVM_RISCV_H
 #define QEMU_KVM_RISCV_H
 
+void kvm_riscv_init_user_properties(Object *cpu_obj);
 void kvm_riscv_reset_vcpu(RISCVCPU *cpu);
 void kvm_riscv_set_irq(RISCVCPU *cpu, int irq, int level);
 
Signature · 1 line
2.41.0
#7 lore raw

PATCH v3 7/19 target/riscv: read marchid/mimpid in kvm_riscv_init_machine_ids()

In reply to Daniel Henrique Barboza

Allow 'marchid' and 'mimpid' to also be initialized in kvm_riscv_init_machine_ids().

After this change, the handling of mvendorid/marchid/mimpid for the 'host' CPU type will be equal to what we already have for TCG named CPUs, i.e. the user is not able to set these values to a different val than the one that is already preset.

---
 target/riscv/kvm.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
target/riscv/kvm.c +16 -0
index 37f0f70794..cd2974c663 100644
@@ -378,6 +378,22 @@ static void kvm_riscv_init_machine_ids(RISCVCPU *cpu, KVMScratchCPU *kvmcpu)
     if (ret != 0) {
         error_report("Unable to retrieve mvendorid from host, error %d", ret);
     }
+
+    reg.id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
+                              KVM_REG_RISCV_CONFIG_REG(marchid));
+    reg.addr = (uint64_t)&cpu->cfg.marchid;
+    ret = ioctl(kvmcpu->cpufd, KVM_GET_ONE_REG, &reg);
+    if (ret != 0) {
+        error_report("Unable to retrieve marchid from host, error %d", ret);
+    }
+
+    reg.id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
+                              KVM_REG_RISCV_CONFIG_REG(mimpid));
+    reg.addr = (uint64_t)&cpu->cfg.mimpid;
+    ret = ioctl(kvmcpu->cpufd, KVM_GET_ONE_REG, &reg);
+    if (ret != 0) {
+        error_report("Unable to retrieve mimpid from host, error %d", ret);
+    }
 }
 
 void kvm_riscv_init_user_properties(Object *cpu_obj)
Signature · 1 line
2.41.0
#8 lore raw

PATCH v3 8/19 target/riscv: handle mvendorid/marchid/mimpid for KVM CPUs

In reply to Daniel Henrique Barboza

After changing user validation for mvendorid/marchid/mimpid to guarantee that the value is validated on user input time, coupled with the work in fetching KVM default values for them by using a scratch CPU, we're certain that the values in cpu->cfg.(mvendorid|marchid|mimpid) are already good to be written back to KVM.

There's no need to write the values back for 'host' type CPUs since the values can't be changed, so let's do that just for generic CPUs.

---
 target/riscv/kvm.c | 31 +++++++++++++++++++++++++++++++
 1 file changed, 31 insertions(+)
target/riscv/kvm.c +31 -0
index cd2974c663..602727cdfd 100644
@@ -495,6 +495,33 @@ void kvm_arch_init_irq_routing(KVMState *s)
 {
 }
 
+static int kvm_vcpu_set_machine_ids(RISCVCPU *cpu, CPUState *cs)
+{
+    CPURISCVState *env = &cpu->env;
+    uint64_t id;
+    int ret;
+
+    id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
+                          KVM_REG_RISCV_CONFIG_REG(mvendorid));
+    ret = kvm_set_one_reg(cs, id, &cpu->cfg.mvendorid);
+    if (ret != 0) {
+        return ret;
+    }
+
+    id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
+                          KVM_REG_RISCV_CONFIG_REG(marchid));
+    ret = kvm_set_one_reg(cs, id, &cpu->cfg.marchid);
+    if (ret != 0) {
+        return ret;
+    }
+
+    id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
+                          KVM_REG_RISCV_CONFIG_REG(mimpid));
+    ret = kvm_set_one_reg(cs, id, &cpu->cfg.mimpid);
+
+    return ret;
+}
+
 int kvm_arch_init_vcpu(CPUState *cs)
 {
     int ret = 0;
@@ -513,6 +540,10 @@ int kvm_arch_init_vcpu(CPUState *cs)
     }
     env->misa_ext = isa;
 
+    if (!object_dynamic_cast(OBJECT(cpu), TYPE_RISCV_CPU_HOST)) {
+        ret = kvm_vcpu_set_machine_ids(cpu, cs);
+    }
+
     return ret;
 }
 
Signature · 1 line
2.41.0
#9 lore raw

PATCH v3 9/19 linux-headers: Update to v6.4-rc1

In reply to Daniel Henrique Barboza

Update to commit ac9a78681b92 ("Linux 6.4-rc1").

---
 include/standard-headers/linux/const.h        |  2 +-
 include/standard-headers/linux/virtio_blk.h   | 18 +++----
 .../standard-headers/linux/virtio_config.h    |  6 +++
 include/standard-headers/linux/virtio_net.h   |  1 +
 linux-headers/asm-arm64/kvm.h                 | 33 ++++++++++++
 linux-headers/asm-riscv/kvm.h                 | 53 ++++++++++++++++++-
 linux-headers/asm-riscv/unistd.h              |  9 ++++
 linux-headers/asm-s390/unistd_32.h            |  1 +
 linux-headers/asm-s390/unistd_64.h            |  1 +
 linux-headers/asm-x86/kvm.h                   |  3 ++
 linux-headers/linux/const.h                   |  2 +-
 linux-headers/linux/kvm.h                     | 12 +++--
 linux-headers/linux/psp-sev.h                 |  7 +++
 linux-headers/linux/userfaultfd.h             | 17 +++++-
 14 files changed, 149 insertions(+), 16 deletions(-)
include/standard-headers/linux/const.h +1 -1
index 5e48987251..1eb84b5087 100644
@@ -28,7 +28,7 @@
 #define _BITUL(x)	(_UL(1) << (x))
 #define _BITULL(x)	(_ULL(1) << (x))
 
-#define __ALIGN_KERNEL(x, a)		__ALIGN_KERNEL_MASK(x, (typeof(x))(a) - 1)
+#define __ALIGN_KERNEL(x, a)		__ALIGN_KERNEL_MASK(x, (__typeof__(x))(a) - 1)
 #define __ALIGN_KERNEL_MASK(x, mask)	(((x) + (mask)) & ~(mask))
 
 #define __KERNEL_DIV_ROUND_UP(n, d) (((n) + (d) - 1) / (d))
include/standard-headers/linux/virtio_blk.h +9 -9
index 7155b1a470..d7be3cf5e4 100644
@@ -138,11 +138,11 @@ struct virtio_blk_config {
 
 	/* Zoned block device characteristics (if VIRTIO_BLK_F_ZONED) */
 	struct virtio_blk_zoned_characteristics {
-		uint32_t zone_sectors;
-		uint32_t max_open_zones;
-		uint32_t max_active_zones;
-		uint32_t max_append_sectors;
-		uint32_t write_granularity;
+		__virtio32 zone_sectors;
+		__virtio32 max_open_zones;
+		__virtio32 max_active_zones;
+		__virtio32 max_append_sectors;
+		__virtio32 write_granularity;
 		uint8_t model;
 		uint8_t unused2[3];
 	} zoned;
@@ -239,11 +239,11 @@ struct virtio_blk_outhdr {
  */
 struct virtio_blk_zone_descriptor {
 	/* Zone capacity */
-	uint64_t z_cap;
+	__virtio64 z_cap;
 	/* The starting sector of the zone */
-	uint64_t z_start;
+	__virtio64 z_start;
 	/* Zone write pointer position in sectors */
-	uint64_t z_wp;
+	__virtio64 z_wp;
 	/* Zone type */
 	uint8_t z_type;
 	/* Zone state */
@@ -252,7 +252,7 @@ struct virtio_blk_zone_descriptor {
 };
 
 struct virtio_blk_zone_report {
-	uint64_t nr_zones;
+	__virtio64 nr_zones;
 	uint8_t reserved[56];
 	struct virtio_blk_zone_descriptor zones[];
 };
include/standard-headers/linux/virtio_config.h +6 -0
index 965ee6ae23..8a7d0dc8b0 100644
@@ -97,6 +97,12 @@
  */
 #define VIRTIO_F_SR_IOV			37
 
+/*
+ * This feature indicates that the driver passes extra data (besides
+ * identifying the virtqueue) in its device notifications.
+ */
+#define VIRTIO_F_NOTIFICATION_DATA	38
+
 /*
  * This feature indicates that the driver can reset a queue individually.
  */
include/standard-headers/linux/virtio_net.h +1 -0
index c0e797067a..2325485f2c 100644
@@ -61,6 +61,7 @@
 #define VIRTIO_NET_F_GUEST_USO6	55	/* Guest can handle USOv6 in. */
 #define VIRTIO_NET_F_HOST_USO	56	/* Host can handle USO in. */
 #define VIRTIO_NET_F_HASH_REPORT  57	/* Supports hash report */
+#define VIRTIO_NET_F_GUEST_HDRLEN  59	/* Guest provides the exact hdr_len value. */
 #define VIRTIO_NET_F_RSS	  60	/* Supports RSS RX steering */
 #define VIRTIO_NET_F_RSC_EXT	  61	/* extended coalescing info */
 #define VIRTIO_NET_F_STANDBY	  62	/* Act as standby for another device
linux-headers/asm-arm64/kvm.h +33 -0
index d7e7bb885e..38e5957526 100644
@@ -198,6 +198,15 @@ struct kvm_arm_copy_mte_tags {
 	__u64 reserved[2];
 };
 
+/*
+ * Counter/Timer offset structure. Describe the virtual/physical offset.
+ * To be used with KVM_ARM_SET_COUNTER_OFFSET.
+ */
+struct kvm_arm_counter_offset {
+	__u64 counter_offset;
+	__u64 reserved;
+};
+
 #define KVM_ARM_TAGS_TO_GUEST		0
 #define KVM_ARM_TAGS_FROM_GUEST		1
 
@@ -363,6 +372,10 @@ enum {
 	KVM_REG_ARM_VENDOR_HYP_BIT_PTP		= 1,
 };
 
+/* Device Control API on vm fd */
+#define KVM_ARM_VM_SMCCC_CTRL		0
+#define   KVM_ARM_VM_SMCCC_FILTER	0
+
 /* Device Control API: ARM VGIC */
 #define KVM_DEV_ARM_VGIC_GRP_ADDR	0
 #define KVM_DEV_ARM_VGIC_GRP_DIST_REGS	1
@@ -402,6 +415,8 @@ enum {
 #define KVM_ARM_VCPU_TIMER_CTRL		1
 #define   KVM_ARM_VCPU_TIMER_IRQ_VTIMER		0
 #define   KVM_ARM_VCPU_TIMER_IRQ_PTIMER		1
+#define   KVM_ARM_VCPU_TIMER_IRQ_HVTIMER	2
+#define   KVM_ARM_VCPU_TIMER_IRQ_HPTIMER	3
 #define KVM_ARM_VCPU_PVTIME_CTRL	2
 #define   KVM_ARM_VCPU_PVTIME_IPA	0
 
@@ -458,6 +473,24 @@ enum {
 /* run->fail_entry.hardware_entry_failure_reason codes. */
 #define KVM_EXIT_FAIL_ENTRY_CPU_UNSUPPORTED	(1ULL << 0)
 
+enum kvm_smccc_filter_action {
+	KVM_SMCCC_FILTER_HANDLE = 0,
+	KVM_SMCCC_FILTER_DENY,
+	KVM_SMCCC_FILTER_FWD_TO_USER,
+
+};
+
+struct kvm_smccc_filter {
+	__u32 base;
+	__u32 nr_functions;
+	__u8 action;
+	__u8 pad[15];
+};
+
+/* arm64-specific KVM_EXIT_HYPERCALL flags */
+#define KVM_HYPERCALL_EXIT_SMC		(1U << 0)
+#define KVM_HYPERCALL_EXIT_16BIT	(1U << 1)
+
 #endif
 
 #endif /* __ARM_KVM_H__ */
linux-headers/asm-riscv/kvm.h +52 -1

106 lines not on this page · raw message

linux-headers/asm-riscv/unistd.h +9 -0
index 73d7cdd2ec..950ab3fd44 100644
@@ -43,3 +43,12 @@
 #define __NR_riscv_flush_icache (__NR_arch_specific_syscall + 15)
 #endif
 __SYSCALL(__NR_riscv_flush_icache, sys_riscv_flush_icache)
+
+/*
+ * Allows userspace to query the kernel for CPU architecture and
+ * microarchitecture details across a given set of CPUs.
+ */
+#ifndef __NR_riscv_hwprobe
+#define __NR_riscv_hwprobe (__NR_arch_specific_syscall + 14)
+#endif
+__SYSCALL(__NR_riscv_hwprobe, sys_riscv_hwprobe)
linux-headers/asm-s390/unistd_32.h +1 -0
index 8e644d65f5..800f3adb20 100644
@@ -419,6 +419,7 @@
 #define __NR_landlock_create_ruleset 444
 #define __NR_landlock_add_rule 445
 #define __NR_landlock_restrict_self 446
+#define __NR_memfd_secret 447
 #define __NR_process_mrelease 448
 #define __NR_futex_waitv 449
 #define __NR_set_mempolicy_home_node 450
linux-headers/asm-s390/unistd_64.h +1 -0
index 51da542fec..399a605901 100644
@@ -367,6 +367,7 @@
 #define __NR_landlock_create_ruleset 444
 #define __NR_landlock_add_rule 445
 #define __NR_landlock_restrict_self 446
+#define __NR_memfd_secret 447
 #define __NR_process_mrelease 448
 #define __NR_futex_waitv 449
 #define __NR_set_mempolicy_home_node 450
linux-headers/asm-x86/kvm.h +3 -0
index 2937e7bf69..2b3a8f7bd2 100644
@@ -557,4 +557,7 @@ struct kvm_pmu_event_filter {
 #define KVM_VCPU_TSC_CTRL 0 /* control group for the timestamp counter (TSC) */
 #define   KVM_VCPU_TSC_OFFSET 0 /* attribute for the TSC offset */
 
+/* x86-specific KVM_EXIT_HYPERCALL flags. */
+#define KVM_EXIT_HYPERCALL_LONG_MODE	BIT(0)
+
 #endif /* _ASM_X86_KVM_H */
linux-headers/linux/const.h +1 -1
index 5e48987251..1eb84b5087 100644
@@ -28,7 +28,7 @@
 #define _BITUL(x)	(_UL(1) << (x))
 #define _BITULL(x)	(_ULL(1) << (x))
 
-#define __ALIGN_KERNEL(x, a)		__ALIGN_KERNEL_MASK(x, (typeof(x))(a) - 1)
+#define __ALIGN_KERNEL(x, a)		__ALIGN_KERNEL_MASK(x, (__typeof__(x))(a) - 1)
 #define __ALIGN_KERNEL_MASK(x, mask)	(((x) + (mask)) & ~(mask))
 
 #define __KERNEL_DIV_ROUND_UP(n, d) (((n) + (d) - 1) / (d))
linux-headers/linux/kvm.h +9 -3
index 599de3c6e3..65b145b317 100644
@@ -341,8 +341,11 @@ struct kvm_run {
 			__u64 nr;
 			__u64 args[6];
 			__u64 ret;
-			__u32 longmode;
-			__u32 pad;
+
+			union {
+				__u32 longmode;
+				__u64 flags;
+			};
 		} hypercall;
 		/* KVM_EXIT_TPR_ACCESS */
 		struct {
@@ -1182,6 +1185,7 @@ struct kvm_ppc_resize_hpt {
 #define KVM_CAP_S390_PROTECTED_ASYNC_DISABLE 224
 #define KVM_CAP_DIRTY_LOG_RING_WITH_BITMAP 225
 #define KVM_CAP_PMU_EVENT_MASKED_EVENTS 226
+#define KVM_CAP_COUNTER_OFFSET 227
 
 #ifdef KVM_CAP_IRQ_ROUTING
 
@@ -1449,7 +1453,7 @@ struct kvm_vfio_spapr_tce {
 #define KVM_CREATE_VCPU           _IO(KVMIO,   0x41)
 #define KVM_GET_DIRTY_LOG         _IOW(KVMIO,  0x42, struct kvm_dirty_log)
 #define KVM_SET_NR_MMU_PAGES      _IO(KVMIO,   0x44)
-#define KVM_GET_NR_MMU_PAGES      _IO(KVMIO,   0x45)
+#define KVM_GET_NR_MMU_PAGES      _IO(KVMIO,   0x45)  /* deprecated */
 #define KVM_SET_USER_MEMORY_REGION _IOW(KVMIO, 0x46, \
 					struct kvm_userspace_memory_region)
 #define KVM_SET_TSS_ADDR          _IO(KVMIO,   0x47)
@@ -1541,6 +1545,8 @@ struct kvm_s390_ucas_mapping {
 #define KVM_SET_PMU_EVENT_FILTER  _IOW(KVMIO,  0xb2, struct kvm_pmu_event_filter)
 #define KVM_PPC_SVM_OFF		  _IO(KVMIO,  0xb3)
 #define KVM_ARM_MTE_COPY_TAGS	  _IOR(KVMIO,  0xb4, struct kvm_arm_copy_mte_tags)
+/* Available with KVM_CAP_COUNTER_OFFSET */
+#define KVM_ARM_SET_COUNTER_OFFSET _IOW(KVMIO,  0xb5, struct kvm_arm_counter_offset)
 
 /* ioctl for vm fd */
 #define KVM_CREATE_DEVICE	  _IOWR(KVMIO,  0xe0, struct kvm_create_device)
linux-headers/linux/psp-sev.h +7 -0
index 51d8b3940e..12ccb70099 100644
@@ -36,6 +36,13 @@ enum {
  * SEV Firmware status code
  */
 typedef enum {
+	/*
+	 * This error code is not in the SEV spec. Its purpose is to convey that
+	 * there was an error that prevented the SEV firmware from being called.
+	 * The SEV API error codes are 16 bits, so the -1 value will not overlap
+	 * with possible values from the specification.
+	 */
+	SEV_RET_NO_FW_CALL = -1,
 	SEV_RET_SUCCESS = 0,
 	SEV_RET_INVALID_PLATFORM_STATE,
 	SEV_RET_INVALID_GUEST_STATE,
linux-headers/linux/userfaultfd.h +16 -1
index ba5d0df52f..14e402263a 100644
@@ -38,7 +38,8 @@
 			   UFFD_FEATURE_MINOR_HUGETLBFS |	\
 			   UFFD_FEATURE_MINOR_SHMEM |		\
 			   UFFD_FEATURE_EXACT_ADDRESS |		\
-			   UFFD_FEATURE_WP_HUGETLBFS_SHMEM)
+			   UFFD_FEATURE_WP_HUGETLBFS_SHMEM |	\
+			   UFFD_FEATURE_WP_UNPOPULATED)
 #define UFFD_API_IOCTLS				\
 	((__u64)1 << _UFFDIO_REGISTER |		\
 	 (__u64)1 << _UFFDIO_UNREGISTER |	\
@@ -203,6 +204,12 @@ struct uffdio_api {
 	 *
 	 * UFFD_FEATURE_WP_HUGETLBFS_SHMEM indicates that userfaultfd
 	 * write-protection mode is supported on both shmem and hugetlbfs.
+	 *
+	 * UFFD_FEATURE_WP_UNPOPULATED indicates that userfaultfd
+	 * write-protection mode will always apply to unpopulated pages
+	 * (i.e. empty ptes).  This will be the default behavior for shmem
+	 * & hugetlbfs, so this flag only affects anonymous memory behavior
+	 * when userfault write-protection mode is registered.
 	 */
 #define UFFD_FEATURE_PAGEFAULT_FLAG_WP		(1<<0)
 #define UFFD_FEATURE_EVENT_FORK			(1<<1)
@@ -217,6 +224,7 @@ struct uffdio_api {
 #define UFFD_FEATURE_MINOR_SHMEM		(1<<10)
 #define UFFD_FEATURE_EXACT_ADDRESS		(1<<11)
 #define UFFD_FEATURE_WP_HUGETLBFS_SHMEM		(1<<12)
+#define UFFD_FEATURE_WP_UNPOPULATED		(1<<13)
 	__u64 features;
 
 	__u64 ioctls;
@@ -297,6 +305,13 @@ struct uffdio_writeprotect {
 struct uffdio_continue {
 	struct uffdio_range range;
 #define UFFDIO_CONTINUE_MODE_DONTWAKE		((__u64)1<<0)
+	/*
+	 * UFFDIO_CONTINUE_MODE_WP will map the page write protected on
+	 * the fly.  UFFDIO_CONTINUE_MODE_WP is available only if the
+	 * write protected ioctl is implemented for the range
+	 * according to the uffdio_register.ioctls.
+	 */
+#define UFFDIO_CONTINUE_MODE_WP			((__u64)1<<1)
 	__u64 mode;
 
 	/*
Signature · 1 line
2.41.0
#10 lore raw

PATCH v3 10/19 target/riscv/kvm.c: init 'misa_ext_mask' with scratch CPU

In reply to Daniel Henrique Barboza

At this moment we're retrieving env->misa_ext during kvm_arch_init_cpu(), leaving env->misa_ext_mask behind.

We want to set env->misa_ext_mask, and we want to set it as early as possible. The reason is that we're going to use it in the validation process of the KVM MISA properties we're going to add next. Setting it during arch_init_cpu() is too late for user validation.

Move the code to a new helper that is going to be called during init() time, via kvm_riscv_init_user_properties(), like we're already doing for the machine ID properties. Set both misa_ext and misa_ext_mask to the same value retrieved by the 'isa' config reg.

---
 target/riscv/kvm.c | 34 +++++++++++++++++++++++-----------
 1 file changed, 23 insertions(+), 11 deletions(-)
target/riscv/kvm.c +23 -11
index 602727cdfd..4d0808cb9a 100644
@@ -396,6 +396,28 @@ static void kvm_riscv_init_machine_ids(RISCVCPU *cpu, KVMScratchCPU *kvmcpu)
     }
 }
 
+static void kvm_riscv_init_misa_ext_mask(RISCVCPU *cpu,
+                                         KVMScratchCPU *kvmcpu)
+{
+    CPURISCVState *env = &cpu->env;
+    struct kvm_one_reg reg;
+    int ret;
+
+    reg.id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
+                              KVM_REG_RISCV_CONFIG_REG(isa));
+    reg.addr = (uint64_t)&env->misa_ext_mask;
+    ret = ioctl(kvmcpu->cpufd, KVM_GET_ONE_REG, &reg);
+
+    if (ret) {
+        error_report("Unable to fetch ISA register from KVM, "
+                     "error %d", ret);
+        kvm_riscv_destroy_scratch_vcpu(kvmcpu);
+        exit(EXIT_FAILURE);
+    }
+
+    env->misa_ext = env->misa_ext_mask;
+}
+
 void kvm_riscv_init_user_properties(Object *cpu_obj)
 {
     RISCVCPU *cpu = RISCV_CPU(cpu_obj);
@@ -406,6 +428,7 @@ void kvm_riscv_init_user_properties(Object *cpu_obj)
     }
 
     kvm_riscv_init_machine_ids(cpu, &kvmcpu);
+    kvm_riscv_init_misa_ext_mask(cpu, &kvmcpu);
 
     kvm_riscv_destroy_scratch_vcpu(&kvmcpu);
 }
@@ -525,21 +548,10 @@ static int kvm_vcpu_set_machine_ids(RISCVCPU *cpu, CPUState *cs)
 int kvm_arch_init_vcpu(CPUState *cs)
 {
     int ret = 0;
-    target_ulong isa;
     RISCVCPU *cpu = RISCV_CPU(cs);
-    CPURISCVState *env = &cpu->env;
-    uint64_t id;
 
     qemu_add_vm_change_state_handler(kvm_riscv_vm_state_change, cs);
 
-    id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
-                          KVM_REG_RISCV_CONFIG_REG(isa));
-    ret = kvm_get_one_reg(cs, id, &isa);
-    if (ret) {
-        return ret;
-    }
-    env->misa_ext = isa;
-
     if (!object_dynamic_cast(OBJECT(cpu), TYPE_RISCV_CPU_HOST)) {
         ret = kvm_vcpu_set_machine_ids(cpu, cs);
     }
Signature · 1 line
2.41.0
#11 lore raw

PATCH v3 11/19 target/riscv/cpu: add misa_ext_info_arr[]

In reply to Daniel Henrique Barboza

Next patch will add KVM specific user properties for both MISA and multi-letter extensions. For MISA extensions we want to make use of what is already available in misa_ext_cfgs[] to avoid code repetition.

misa_ext_info_arr[] array will hold name and description for each MISA extension that misa_ext_cfgs[] is declaring. We'll then use this new array in KVM code to avoid duplicating strings.

There's nothing holding us back from doing the same with multi-letter extensions. For now doing just with MISA extensions is enough.

---
 target/riscv/cpu.c | 60 ++++++++++++++++++++++++++--------------------
 target/riscv/cpu.h | 11 ++++++++-
 2 files changed, 44 insertions(+), 27 deletions(-)
target/riscv/cpu.c +34 -26
index 641bec3573..0e5d8b05a2 100644
@@ -1566,33 +1566,41 @@ static void cpu_get_misa_ext_cfg(Object *obj, Visitor *v, const char *name,
     visit_type_bool(v, name, &value, errp);
 }
 
+const MISAExtInfo misa_ext_info_arr[] = {
+    [RVA] = {"a", "Atomic instructions"},
+    [RVC] = {"c", "Compressed instructions"},
+    [RVD] = {"d", "Double-precision float point"},
+    [RVF] = {"f", "Single-precision float point"},
+    [RVI] = {"i", "Base integer instruction set"},
+    [RVE] = {"e", "Base integer instruction set (embedded)"},
+    [RVM] = {"m", "Integer multiplication and division"},
+    [RVS] = {"s", "Supervisor-level instructions"},
+    [RVU] = {"u", "User-level instructions"},
+    [RVH] = {"h", "Hypervisor"},
+    [RVJ] = {"x-j", "Dynamic translated languages"},
+    [RVV] = {"v", "Vector operations"},
+    [RVG] = {"g", "General purpose (IMAFD_Zicsr_Zifencei)"},
+};
+
+#define MISA_CFG(_bit, _enabled) \
+    {.name = misa_ext_info_arr[_bit].name, \
+     .description = misa_ext_info_arr[_bit].description, \
+     .misa_bit = _bit, .enabled = _enabled}
+
 static const RISCVCPUMisaExtConfig misa_ext_cfgs[] = {
-    {.name = "a", .description = "Atomic instructions",
-     .misa_bit = RVA, .enabled = true},
-    {.name = "c", .description = "Compressed instructions",
-     .misa_bit = RVC, .enabled = true},
-    {.name = "d", .description = "Double-precision float point",
-     .misa_bit = RVD, .enabled = true},
-    {.name = "f", .description = "Single-precision float point",
-     .misa_bit = RVF, .enabled = true},
-    {.name = "i", .description = "Base integer instruction set",
-     .misa_bit = RVI, .enabled = true},
-    {.name = "e", .description = "Base integer instruction set (embedded)",
-     .misa_bit = RVE, .enabled = false},
-    {.name = "m", .description = "Integer multiplication and division",
-     .misa_bit = RVM, .enabled = true},
-    {.name = "s", .description = "Supervisor-level instructions",
-     .misa_bit = RVS, .enabled = true},
-    {.name = "u", .description = "User-level instructions",
-     .misa_bit = RVU, .enabled = true},
-    {.name = "h", .description = "Hypervisor",
-     .misa_bit = RVH, .enabled = true},
-    {.name = "x-j", .description = "Dynamic translated languages",
-     .misa_bit = RVJ, .enabled = false},
-    {.name = "v", .description = "Vector operations",
-     .misa_bit = RVV, .enabled = false},
-    {.name = "g", .description = "General purpose (IMAFD_Zicsr_Zifencei)",
-     .misa_bit = RVG, .enabled = false},
+    MISA_CFG(RVA, true),
+    MISA_CFG(RVC, true),
+    MISA_CFG(RVD, true),
+    MISA_CFG(RVF, true),
+    MISA_CFG(RVI, true),
+    MISA_CFG(RVE, false),
+    MISA_CFG(RVM, true),
+    MISA_CFG(RVS, true),
+    MISA_CFG(RVU, true),
+    MISA_CFG(RVH, true),
+    MISA_CFG(RVJ, false),
+    MISA_CFG(RVV, false),
+    MISA_CFG(RVG, false),
 };
 
 static void riscv_cpu_add_misa_properties(Object *cpu_obj)
target/riscv/cpu.h +10 -1
index e3e08d315f..3e08c8da51 100644
@@ -41,7 +41,10 @@
 
 #define RV(x) ((target_ulong)1 << (x - 'A'))
 
-/* Consider updating misa_ext_cfgs[] when adding new MISA bits here */
+/*
+ * Consider updating misa_ext_info_arr[] and misa_ext_cfgs[]
+ * when adding new MISA bits here.
+ */
 #define RVI RV('I')
 #define RVE RV('E') /* E and I are mutually exclusive */
 #define RVM RV('M')
@@ -56,6 +59,12 @@
 #define RVJ RV('J')
 #define RVG RV('G')
 
+typedef struct misa_ext_info {
+    const char *name;
+    const char *description;
+} MISAExtInfo;
+
+extern const MISAExtInfo misa_ext_info_arr[];
 
 /* Privileged specification version */
 enum {
Signature · 1 line
2.41.0
Reviewed-by on v3 11/19
#12 lore raw

Re: [PATCH v3 11/19] target/riscv/cpu: add misa_ext_info_arr[]

Quoted 122 lines · Daniel Henrique Barboza

Next patch will add KVM specific user properties for both MISA and multi-letter extensions. For MISA extensions we want to make use of what is already available in misa_ext_cfgs[] to avoid code repetition.

misa_ext_info_arr[] array will hold name and description for each MISA extension that misa_ext_cfgs[] is declaring. We'll then use this new array in KVM code to avoid duplicating strings.

There's nothing holding us back from doing the same with multi-letter extensions. For now doing just with MISA extensions is enough.

---
 target/riscv/cpu.c | 60 ++++++++++++++++++++++++++--------------------
 target/riscv/cpu.h | 11 ++++++++-
 2 files changed, 44 insertions(+), 27 deletions(-)
target/riscv/cpu.c +34 -26
index 641bec3573..0e5d8b05a2 100644
@@ -1566,33 +1566,41 @@ static void cpu_get_misa_ext_cfg(Object *obj, Visitor *v, const char *name,
     visit_type_bool(v, name, &value, errp);
 }
 
+const MISAExtInfo misa_ext_info_arr[] = {
+    [RVA] = {"a", "Atomic instructions"},
+    [RVC] = {"c", "Compressed instructions"},
+    [RVD] = {"d", "Double-precision float point"},
+    [RVF] = {"f", "Single-precision float point"},
+    [RVI] = {"i", "Base integer instruction set"},
+    [RVE] = {"e", "Base integer instruction set (embedded)"},
+    [RVM] = {"m", "Integer multiplication and division"},
+    [RVS] = {"s", "Supervisor-level instructions"},
+    [RVU] = {"u", "User-level instructions"},
+    [RVH] = {"h", "Hypervisor"},
+    [RVJ] = {"x-j", "Dynamic translated languages"},
+    [RVV] = {"v", "Vector operations"},
+    [RVG] = {"g", "General purpose (IMAFD_Zicsr_Zifencei)"},
+};
+
+#define MISA_CFG(_bit, _enabled) \
+    {.name = misa_ext_info_arr[_bit].name, \
+     .description = misa_ext_info_arr[_bit].description, \
+     .misa_bit = _bit, .enabled = _enabled}
+
 static const RISCVCPUMisaExtConfig misa_ext_cfgs[] = {
-    {.name = "a", .description = "Atomic instructions",
-     .misa_bit = RVA, .enabled = true},
-    {.name = "c", .description = "Compressed instructions",
-     .misa_bit = RVC, .enabled = true},
-    {.name = "d", .description = "Double-precision float point",
-     .misa_bit = RVD, .enabled = true},
-    {.name = "f", .description = "Single-precision float point",
-     .misa_bit = RVF, .enabled = true},
-    {.name = "i", .description = "Base integer instruction set",
-     .misa_bit = RVI, .enabled = true},
-    {.name = "e", .description = "Base integer instruction set (embedded)",
-     .misa_bit = RVE, .enabled = false},
-    {.name = "m", .description = "Integer multiplication and division",
-     .misa_bit = RVM, .enabled = true},
-    {.name = "s", .description = "Supervisor-level instructions",
-     .misa_bit = RVS, .enabled = true},
-    {.name = "u", .description = "User-level instructions",
-     .misa_bit = RVU, .enabled = true},
-    {.name = "h", .description = "Hypervisor",
-     .misa_bit = RVH, .enabled = true},
-    {.name = "x-j", .description = "Dynamic translated languages",
-     .misa_bit = RVJ, .enabled = false},
-    {.name = "v", .description = "Vector operations",
-     .misa_bit = RVV, .enabled = false},
-    {.name = "g", .description = "General purpose (IMAFD_Zicsr_Zifencei)",
-     .misa_bit = RVG, .enabled = false},
+    MISA_CFG(RVA, true),
+    MISA_CFG(RVC, true),
+    MISA_CFG(RVD, true),
+    MISA_CFG(RVF, true),
+    MISA_CFG(RVI, true),
+    MISA_CFG(RVE, false),
+    MISA_CFG(RVM, true),
+    MISA_CFG(RVS, true),
+    MISA_CFG(RVU, true),
+    MISA_CFG(RVH, true),
+    MISA_CFG(RVJ, false),
+    MISA_CFG(RVV, false),
+    MISA_CFG(RVG, false),
 };
 
 static void riscv_cpu_add_misa_properties(Object *cpu_obj)
target/riscv/cpu.h +10 -1
index e3e08d315f..3e08c8da51 100644
@@ -41,7 +41,10 @@
 
 #define RV(x) ((target_ulong)1 << (x - 'A'))
 
-/* Consider updating misa_ext_cfgs[] when adding new MISA bits here */
+/*
+ * Consider updating misa_ext_info_arr[] and misa_ext_cfgs[]
+ * when adding new MISA bits here.
+ */
 #define RVI RV('I')
 #define RVE RV('E') /* E and I are mutually exclusive */
 #define RVM RV('M')
@@ -56,6 +59,12 @@
 #define RVJ RV('J')
 #define RVG RV('G')
 
+typedef struct misa_ext_info {
+    const char *name;
+    const char *description;
+} MISAExtInfo;
+
+extern const MISAExtInfo misa_ext_info_arr[];
 
 /* Privileged specification version */
 enum {
Signature · 1 line
2.41.0
#13 lore raw

PATCH v3 12/19 target/riscv: add KVM specific MISA properties

In reply to Daniel Henrique Barboza

Using all TCG user properties in KVM is tricky. First because KVM supports only a small subset of what TCG provides, so most of the cpu->cfg flags do nothing for KVM.

Second, and more important, we don't have a way of telling if any given value is an user input or not. For TCG this has a small impact since we just validating everything and error out if needed. But for KVM it would be good to know if a given value was set by the user or if it's a value already provided by KVM. Otherwise we don't know how to handle failed kvm_set_one_regs() when writing the configurations back.

These characteristics make it overly complicated to use the same user facing flags for both KVM and TCG. A simpler approach is to create KVM specific properties that have specialized logic, forking KVM and TCG use cases for those cases only. Fully separating KVM/TCG properties is unneeded at this point - in fact we want the user experience to be as equal as possible, regardless of the acceleration chosen.

We'll start this fork with the MISA properties, adding the MISA bits that the KVM driver currently supports. A new KVMCPUConfig type is introduced. It'll hold general information about an extension. For MISA extensions we're going to use the newly created misa_ext_infos[] to populate their name and description. 'offset' holds the MISA bit (RVA, RVC, ...). We're calling it 'offset' instead of 'misa_bit' because this same KVMCPUConfig struct will be used to multi-letter extensions later on.

This new type also holds a 'user_set' flag. This flag will be set when the user set an option that's different than what is already configured in the host, requiring KVM intervention to write the regs back during kvm_arch_init_vcpu(). Similar mechanics will be implemented for multi-letter extensions as well.

There is no need to duplicate more code than necessary, so we're going to use the existing kvm_riscv_init_user_properties() to add the KVM specific properties. Any code that is adding a TCG user prop is then changed slightly to verify first if there's a KVM prop with the same name already added.

---
 target/riscv/cpu.c | 10 ++++++
 target/riscv/kvm.c | 78 ++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 88 insertions(+)
target/riscv/cpu.c +10 -0
index 0e5d8b05a2..5a50cb1d0b 100644
@@ -1610,6 +1610,11 @@ static void riscv_cpu_add_misa_properties(Object *cpu_obj)
     for (i = 0; i < ARRAY_SIZE(misa_ext_cfgs); i++) {
         const RISCVCPUMisaExtConfig *misa_cfg = &misa_ext_cfgs[i];
 
+        /* Check if KVM didn't create the property already */
+        if (object_property_find(cpu_obj, misa_cfg->name)) {
+            continue;
+        }
+
         object_property_add(cpu_obj, misa_cfg->name, "bool",
                             cpu_get_misa_ext_cfg,
                             cpu_set_misa_ext_cfg,
@@ -1733,6 +1738,11 @@ static void riscv_cpu_add_user_properties(Object *obj)
     riscv_cpu_add_misa_properties(obj);
 
     for (prop = riscv_cpu_extensions; prop && prop->name; prop++) {
+        /* Check if KVM didn't create the property already */
+        if (object_property_find(obj, prop->name)) {
+            continue;
+        }
+
         qdev_property_add_static(dev, prop);
     }
 
target/riscv/kvm.c +78 -0

98 lines not on this page · raw message

Signature · 1 line
2.41.0
#14 lore raw

Re: [PATCH v3 12/19] target/riscv: add KVM specific MISA properties

On Thu, Jun 22, 2023 at 10:56:53AM -0300, Daniel Henrique Barboza wrote:
...

Quoted 14 lines
target/riscv/kvm.c +14 -0
+#define KVM_MISA_CFG(_bit, _reg_id) \
+    {.offset = _bit, .kvm_reg_id = _reg_id}
+
+/* KVM ISA extensions */
+static KVMCPUConfig kvm_misa_ext_cfgs[] = {
+    KVM_MISA_CFG(RVA, KVM_RISCV_ISA_EXT_A),
+    KVM_MISA_CFG(RVC, KVM_RISCV_ISA_EXT_C),
+    KVM_MISA_CFG(RVD, KVM_RISCV_ISA_EXT_D),
+    KVM_MISA_CFG(RVF, KVM_RISCV_ISA_EXT_F),
+    KVM_MISA_CFG(RVH, KVM_RISCV_ISA_EXT_H),
+    KVM_MISA_CFG(RVI, KVM_RISCV_ISA_EXT_I),
+    KVM_MISA_CFG(RVM, KVM_RISCV_ISA_EXT_M),
+};
+

...

target/riscv/kvm.c static uint64_t kvm_riscv_reg_id(CPURISCVState *env, uint64_t type, · line 174

Quoted 10 lines
target/riscv/kvm.c +10 -0
+static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
+{
+    int i;
+
+    for (i = 0; i < ARRAY_SIZE(kvm_misa_ext_cfgs); i++) {
+        KVMCPUConfig *misa_cfg = &kvm_misa_ext_cfgs[i];
+        int bit = misa_cfg->offset;
+
+        misa_cfg->name = misa_ext_info_arr[bit].name;
+        misa_cfg->description = misa_ext_info_arr[bit].description;

I'd prefer these be set by KVM_MISA_CFG(), since we can. No need to wait until runtime if we can do it at compile-time.

Thanks,
drew

Quoted 24 lines
target/riscv/kvm.c +11 -0
+
+        object_property_add(cpu_obj, misa_cfg->name, "bool",
+                            NULL,
+                            kvm_cpu_set_misa_ext_cfg,
+                            NULL, misa_cfg);
+        object_property_set_description(cpu_obj, misa_cfg->name,
+                                        misa_cfg->description);
+    }
+}
+
 static int kvm_riscv_get_regs_core(CPUState *cs)
 {
     int ret = 0;
@@ -427,6 +504,7 @@ void kvm_riscv_init_user_properties(Object *cpu_obj)
         return;
     }
 
+    kvm_riscv_add_cpu_user_properties(cpu_obj);
     kvm_riscv_init_machine_ids(cpu, &kvmcpu);
     kvm_riscv_init_misa_ext_mask(cpu, &kvmcpu);
 
Signature · 1 line
2.41.0
#15 lore raw

Re: [PATCH v3 12/19] target/riscv: add KVM specific MISA properties

Quoted 17 lines · Andrew Jones

On Thu, Jun 22, 2023 at 10:56:53AM -0300, Daniel Henrique Barboza wrote:
...

Quoted 14 lines
target/riscv/kvm.c +14 -0
+#define KVM_MISA_CFG(_bit, _reg_id) \
+    {.offset = _bit, .kvm_reg_id = _reg_id}
+
+/* KVM ISA extensions */
+static KVMCPUConfig kvm_misa_ext_cfgs[] = {
+    KVM_MISA_CFG(RVA, KVM_RISCV_ISA_EXT_A),
+    KVM_MISA_CFG(RVC, KVM_RISCV_ISA_EXT_C),
+    KVM_MISA_CFG(RVD, KVM_RISCV_ISA_EXT_D),
+    KVM_MISA_CFG(RVF, KVM_RISCV_ISA_EXT_F),
+    KVM_MISA_CFG(RVH, KVM_RISCV_ISA_EXT_H),
+    KVM_MISA_CFG(RVI, KVM_RISCV_ISA_EXT_I),
+    KVM_MISA_CFG(RVM, KVM_RISCV_ISA_EXT_M),
+};
+

...

Quoted 13 lines

target/riscv/kvm.c static uint64_t kvm_riscv_reg_id(CPURISCVState *env, uint64_t type, · line 174

Quoted 10 lines
target/riscv/kvm.c +10 -0
+static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
+{
+    int i;
+
+    for (i = 0; i < ARRAY_SIZE(kvm_misa_ext_cfgs); i++) {
+        KVMCPUConfig *misa_cfg = &kvm_misa_ext_cfgs[i];
+        int bit = misa_cfg->offset;
+
+        misa_cfg->name = misa_ext_info_arr[bit].name;
+        misa_cfg->description = misa_ext_info_arr[bit].description;

I'd prefer these be set by KVM_MISA_CFG(), since we can. No need to wait until runtime if we can do it at compile-time.

The compiler will complain about "error: initializer element is not constant" and the build will fail. This happens because, apparently, the compiler doesn't see the imported array as a constant, regardless of the 'const' type.

If there's a trick that allow us to set these inside that macro I'm all ears. I tried a bit, failed, and ended up adding these 2 extra lines. Thanks,

Daniel

Quoted 28 lines

Thanks,
drew

Quoted 24 lines
target/riscv/kvm.c +11 -0
+
+        object_property_add(cpu_obj, misa_cfg->name, "bool",
+                            NULL,
+                            kvm_cpu_set_misa_ext_cfg,
+                            NULL, misa_cfg);
+        object_property_set_description(cpu_obj, misa_cfg->name,
+                                        misa_cfg->description);
+    }
+}
+
  static int kvm_riscv_get_regs_core(CPUState *cs)
  {
      int ret = 0;
@@ -427,6 +504,7 @@ void kvm_riscv_init_user_properties(Object *cpu_obj)
          return;
      }
 
+    kvm_riscv_add_cpu_user_properties(cpu_obj);
      kvm_riscv_init_machine_ids(cpu, &kvmcpu);
      kvm_riscv_init_misa_ext_mask(cpu, &kvmcpu);
 
Signature · 1 line
2.41.0
#16 lore raw

Re: [PATCH v3 12/19] target/riscv: add KVM specific MISA properties

Quoted 34 lines · Daniel Henrique Barboza
Quoted 17 lines · Andrew Jones

On Thu, Jun 22, 2023 at 10:56:53AM -0300, Daniel Henrique Barboza wrote:
...

Quoted 14 lines
target/riscv/kvm.c +14 -0
+#define KVM_MISA_CFG(_bit, _reg_id) \
+    {.offset = _bit, .kvm_reg_id = _reg_id}
+
+/* KVM ISA extensions */
+static KVMCPUConfig kvm_misa_ext_cfgs[] = {
+    KVM_MISA_CFG(RVA, KVM_RISCV_ISA_EXT_A),
+    KVM_MISA_CFG(RVC, KVM_RISCV_ISA_EXT_C),
+    KVM_MISA_CFG(RVD, KVM_RISCV_ISA_EXT_D),
+    KVM_MISA_CFG(RVF, KVM_RISCV_ISA_EXT_F),
+    KVM_MISA_CFG(RVH, KVM_RISCV_ISA_EXT_H),
+    KVM_MISA_CFG(RVI, KVM_RISCV_ISA_EXT_I),
+    KVM_MISA_CFG(RVM, KVM_RISCV_ISA_EXT_M),
+};
+

...

Quoted 13 lines

target/riscv/kvm.c static uint64_t kvm_riscv_reg_id(CPURISCVState *env, uint64_t type, · line 174

Quoted 10 lines
target/riscv/kvm.c +10 -0
+static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
+{
+    int i;
+
+    for (i = 0; i < ARRAY_SIZE(kvm_misa_ext_cfgs); i++) {
+        KVMCPUConfig *misa_cfg = &kvm_misa_ext_cfgs[i];
+        int bit = misa_cfg->offset;
+
+        misa_cfg->name = misa_ext_info_arr[bit].name;
+        misa_cfg->description = misa_ext_info_arr[bit].description;

I'd prefer these be set by KVM_MISA_CFG(), since we can. No need to wait until runtime if we can do it at compile-time.

Quoted 4 lines

The compiler will complain about "error: initializer element is not constant" and the build will fail. This happens because, apparently, the compiler doesn't see the imported array as a constant, regardless of the 'const' type.

You're right. Initialization the way I suggested only works when everything is in the same source file. That's a pity. So we can either manage it they way you've done here or the way you did in a previous patch, which was to include the info array in each source file by putting it in the header. I think I prefer the approach you used in this version more though.

Thanks,
drew

#17 lore raw

Re: [PATCH v3 12/19] target/riscv: add KVM specific MISA properties

Quoted 40 lines · Andrew Jones
Quoted 34 lines · Daniel Henrique Barboza
Quoted 17 lines · Andrew Jones

On Thu, Jun 22, 2023 at 10:56:53AM -0300, Daniel Henrique Barboza wrote:
...

Quoted 14 lines
target/riscv/kvm.c +14 -0
+#define KVM_MISA_CFG(_bit, _reg_id) \
+    {.offset = _bit, .kvm_reg_id = _reg_id}
+
+/* KVM ISA extensions */
+static KVMCPUConfig kvm_misa_ext_cfgs[] = {
+    KVM_MISA_CFG(RVA, KVM_RISCV_ISA_EXT_A),
+    KVM_MISA_CFG(RVC, KVM_RISCV_ISA_EXT_C),
+    KVM_MISA_CFG(RVD, KVM_RISCV_ISA_EXT_D),
+    KVM_MISA_CFG(RVF, KVM_RISCV_ISA_EXT_F),
+    KVM_MISA_CFG(RVH, KVM_RISCV_ISA_EXT_H),
+    KVM_MISA_CFG(RVI, KVM_RISCV_ISA_EXT_I),
+    KVM_MISA_CFG(RVM, KVM_RISCV_ISA_EXT_M),
+};
+

...

Quoted 13 lines

target/riscv/kvm.c static uint64_t kvm_riscv_reg_id(CPURISCVState *env, uint64_t type, · line 174

Quoted 10 lines
target/riscv/kvm.c +10 -0
+static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
+{
+    int i;
+
+    for (i = 0; i < ARRAY_SIZE(kvm_misa_ext_cfgs); i++) {
+        KVMCPUConfig *misa_cfg = &kvm_misa_ext_cfgs[i];
+        int bit = misa_cfg->offset;
+
+        misa_cfg->name = misa_ext_info_arr[bit].name;
+        misa_cfg->description = misa_ext_info_arr[bit].description;

I'd prefer these be set by KVM_MISA_CFG(), since we can. No need to wait until runtime if we can do it at compile-time.

Quoted 4 lines

The compiler will complain about "error: initializer element is not constant" and the build will fail. This happens because, apparently, the compiler doesn't see the imported array as a constant, regardless of the 'const' type.

Quoted 6 lines

You're right. Initialization the way I suggested only works when everything is in the same source file. That's a pity. So we can either manage it they way you've done here or the way you did in a previous patch, which was to include the info array in each source file by putting it in the header. I think I prefer the approach you used in this version more though.

Alright, let's keep this version then.

Btw I just noticed that I kept your reviewed-by from the previous version in this patch by accident. Can you please send a r-b for this version (in case you agree, of course)?

Thanks,

Daniel

Quoted 3 lines

Thanks,
drew

Reviewed-by on v3 12/19
#18 lore raw

Re: [PATCH v3 12/19] target/riscv: add KVM specific MISA properties

In reply to Daniel Henrique Barboza
Quoted 39 lines · Daniel Henrique Barboza

Using all TCG user properties in KVM is tricky. First because KVM supports only a small subset of what TCG provides, so most of the cpu->cfg flags do nothing for KVM.

Second, and more important, we don't have a way of telling if any given value is an user input or not. For TCG this has a small impact since we just validating everything and error out if needed. But for KVM it would be good to know if a given value was set by the user or if it's a value already provided by KVM. Otherwise we don't know how to handle failed kvm_set_one_regs() when writing the configurations back.

These characteristics make it overly complicated to use the same user facing flags for both KVM and TCG. A simpler approach is to create KVM specific properties that have specialized logic, forking KVM and TCG use cases for those cases only. Fully separating KVM/TCG properties is unneeded at this point - in fact we want the user experience to be as equal as possible, regardless of the acceleration chosen.

We'll start this fork with the MISA properties, adding the MISA bits that the KVM driver currently supports. A new KVMCPUConfig type is introduced. It'll hold general information about an extension. For MISA extensions we're going to use the newly created misa_ext_infos[] to populate their name and description. 'offset' holds the MISA bit (RVA, RVC, ...). We're calling it 'offset' instead of 'misa_bit' because this same KVMCPUConfig struct will be used to multi-letter extensions later on.

This new type also holds a 'user_set' flag. This flag will be set when the user set an option that's different than what is already configured in the host, requiring KVM intervention to write the regs back during kvm_arch_init_vcpu(). Similar mechanics will be implemented for multi-letter extensions as well.

There is no need to duplicate more code than necessary, so we're going to use the existing kvm_riscv_init_user_properties() to add the KVM specific properties. Any code that is adding a TCG user prop is then changed slightly to verify first if there's a KVM prop with the same name already added.

Quoted 7 lines
---
 target/riscv/cpu.c | 10 ++++++
 target/riscv/kvm.c | 78 ++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 88 insertions(+)
#19 lore raw

PATCH v3 13/19 target/riscv/kvm.c: update KVM MISA bits

In reply to Daniel Henrique Barboza

Our design philosophy with KVM properties can be resumed in two main decisions based on KVM interface availability and what the user wants to do:

- if the user disables an extension that the host KVM module doesn't know about (i.e. it doesn't implement the kvm_get_one_reg() interface), keep booting the CPU. This will avoid users having to deal with issues with older KVM versions while disabling features they don't care;

- for any other case we're going to error out immediately. If the user wants to enable a feature that KVM doesn't know about this a problem that is worth aborting - the user must know that the feature wasn't enabled in the hart. Likewise, if KVM knows about the extension, the user wants to enable/disable it, and we fail to do it so, that's also a problem we can't shrug it off.

In the case of MISA bits we won't even try enabling bits that aren't already available in the host. The ioctl() is so likely to fail that it's not worth trying. This check is already done in the previous patch, in kvm_cpu_set_misa_ext_cfg(), thus we don't need to worry about it now.

In kvm_riscv_update_cpu_misa_ext() we'll go through every potential user option and do as follows:

- if the user didn't set the property or set to the same value of the host, do nothing;

- Disable the given extension in KVM. Error out if anything goes wrong.

---
 target/riscv/kvm.c | 40 ++++++++++++++++++++++++++++++++++++++++
 1 file changed, 40 insertions(+)
target/riscv/kvm.c +40 -0
index fb5c5d4bfc..cb511ea8f1 100644
@@ -162,6 +162,41 @@ static void kvm_cpu_set_misa_ext_cfg(Object *obj, Visitor *v,
                "enabled in the host", misa_ext_cfg->name);
 }
 
+static void kvm_riscv_update_cpu_misa_ext(RISCVCPU *cpu, CPUState *cs)
+{
+    CPURISCVState *env = &cpu->env;
+    uint64_t id, reg;
+    int i, ret;
+
+    for (i = 0; i < ARRAY_SIZE(kvm_misa_ext_cfgs); i++) {
+        KVMCPUConfig *misa_cfg = &kvm_misa_ext_cfgs[i];
+        target_ulong misa_bit = misa_cfg->offset;
+
+        if (!misa_cfg->user_set) {
+            continue;
+        }
+
+        /* If we're here we're going to disable the MISA bit */
+        reg = 0;
+        id = kvm_riscv_reg_id(env, KVM_REG_RISCV_ISA_EXT,
+                              misa_cfg->kvm_reg_id);
+        ret = kvm_set_one_reg(cs, id, &reg);
+        if (ret != 0) {
+            /*
+             * We're not checking for -EINVAL because if the bit is about
+             * to be disabled, it means that it was already enabled by
+             * KVM. We determined that by fetching the 'isa' register
+             * during init() time. Any error at this point is worth
+             * aborting.
+             */
+            error_report("Unable to set KVM reg %s, error %d",
+                         misa_cfg->name, ret);
+            exit(EXIT_FAILURE);
+        }
+        env->misa_ext &= ~misa_bit;
+    }
+}
+
 static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
 {
     int i;
@@ -632,8 +667,13 @@ int kvm_arch_init_vcpu(CPUState *cs)
 
     if (!object_dynamic_cast(OBJECT(cpu), TYPE_RISCV_CPU_HOST)) {
         ret = kvm_vcpu_set_machine_ids(cpu, cs);
+        if (ret != 0) {
+            return ret;
+        }
     }
 
+    kvm_riscv_update_cpu_misa_ext(cpu, cs);
+
     return ret;
 }
 
Signature · 1 line
2.41.0
#20 lore raw

PATCH v3 14/19 target/riscv/kvm.c: add multi-letter extension KVM properties

In reply to Daniel Henrique Barboza

Let's add KVM user properties for the multi-letter extensions that KVM currently supports: zicbom, zicboz, zihintpause, zbb, ssaia, sstc, svinval and svpbmt.

As with MISA extensions, we're using the KVMCPUConfig type to hold information about the state of each extension. However, multi-letter extensions have more cases to cover than MISA extensions, so we're adding an extra 'supported' flag as well. This flag will reflect if a given extension is supported by KVM, i.e. KVM knows how to handle it. This is determined during KVM extension discovery in kvm_riscv_init_multiext_cfg(), where we test for EINVAL errors. Any other error different from EINVAL will cause an abort.

The use of the 'user_set' is similar to what we already do with MISA
extensions: the flag set only if the user is changing the extension state.

The 'supported' flag will be used later on to make an exception for users that are disabling multi-letter extensions that are unknown to KVM.

---
 target/riscv/kvm.c | 119 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 119 insertions(+)
target/riscv/kvm.c +119 -0
index cb511ea8f1..376bbc68b4 100644
@@ -113,6 +113,7 @@ typedef struct KVMCPUConfig {
     target_ulong offset;
     int kvm_reg_id;
     bool user_set;
+    bool supported;
 } KVMCPUConfig;
 
 #define KVM_MISA_CFG(_bit, _reg_id) \
@@ -197,6 +198,81 @@ static void kvm_riscv_update_cpu_misa_ext(RISCVCPU *cpu, CPUState *cs)
     }
 }
 
+#define CPUCFG(_prop) offsetof(struct RISCVCPUConfig, _prop)
+
+#define KVM_EXT_CFG(_name, _prop, _reg_id) \
+    {.name = _name, .offset = CPUCFG(_prop), \
+     .kvm_reg_id = _reg_id}
+
+static KVMCPUConfig kvm_multi_ext_cfgs[] = {
+    KVM_EXT_CFG("zicbom", ext_icbom, KVM_RISCV_ISA_EXT_ZICBOM),
+    KVM_EXT_CFG("zicboz", ext_icboz, KVM_RISCV_ISA_EXT_ZICBOZ),
+    KVM_EXT_CFG("zihintpause", ext_zihintpause, KVM_RISCV_ISA_EXT_ZIHINTPAUSE),
+    KVM_EXT_CFG("zbb", ext_zbb, KVM_RISCV_ISA_EXT_ZBB),
+    KVM_EXT_CFG("ssaia", ext_ssaia, KVM_RISCV_ISA_EXT_SSAIA),
+    KVM_EXT_CFG("sstc", ext_sstc, KVM_RISCV_ISA_EXT_SSTC),
+    KVM_EXT_CFG("svinval", ext_svinval, KVM_RISCV_ISA_EXT_SVINVAL),
+    KVM_EXT_CFG("svpbmt", ext_svpbmt, KVM_RISCV_ISA_EXT_SVPBMT),
+};
+
+static void kvm_cpu_cfg_set(RISCVCPU *cpu, KVMCPUConfig *multi_ext,
+                            uint32_t val)
+{
+    int cpu_cfg_offset = multi_ext->offset;
+    bool *ext_enabled = (void *)&cpu->cfg + cpu_cfg_offset;
+
+    *ext_enabled = val;
+}
+
+static uint32_t kvm_cpu_cfg_get(RISCVCPU *cpu,
+                                KVMCPUConfig *multi_ext)
+{
+    int cpu_cfg_offset = multi_ext->offset;
+    bool *ext_enabled = (void *)&cpu->cfg + cpu_cfg_offset;
+
+    return *ext_enabled;
+}
+
+static void kvm_cpu_set_multi_ext_cfg(Object *obj, Visitor *v,
+                                      const char *name,
+                                      void *opaque, Error **errp)
+{
+    KVMCPUConfig *multi_ext_cfg = opaque;
+    RISCVCPU *cpu = RISCV_CPU(obj);
+    bool value, host_val;
+
+    if (!visit_type_bool(v, name, &value, errp)) {
+        return;
+    }
+
+    host_val = kvm_cpu_cfg_get(cpu, multi_ext_cfg);
+
+    /*
+     * Ignore if the user is setting the same value
+     * as the host.
+     */
+    if (value == host_val) {
+        return;
+    }
+
+    if (!multi_ext_cfg->supported) {
+        /*
+         * Error out if the user is trying to enable an
+         * extension that KVM doesn't support. Ignore
+         * option otherwise.
+         */
+        if (value) {
+            error_setg(errp, "KVM does not support disabling extension %s",
+                       multi_ext_cfg->name);
+        }
+
+        return;
+    }
+
+    multi_ext_cfg->user_set = true;
+    kvm_cpu_cfg_set(cpu, multi_ext_cfg, value);
+}
+
 static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
 {
     int i;
@@ -215,6 +291,15 @@ static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
         object_property_set_description(cpu_obj, misa_cfg->name,
                                         misa_cfg->description);
     }
+
+    for (i = 0; i < ARRAY_SIZE(kvm_multi_ext_cfgs); i++) {
+        KVMCPUConfig *multi_cfg = &kvm_multi_ext_cfgs[i];
+
+        object_property_add(cpu_obj, multi_cfg->name, "bool",
+                            NULL,
+                            kvm_cpu_set_multi_ext_cfg,
+                            NULL, multi_cfg);
+    }
 }
 
 static int kvm_riscv_get_regs_core(CPUState *cs)
@@ -530,6 +615,39 @@ static void kvm_riscv_init_misa_ext_mask(RISCVCPU *cpu,
     env->misa_ext = env->misa_ext_mask;
 }
 
+static void kvm_riscv_init_multiext_cfg(RISCVCPU *cpu, KVMScratchCPU *kvmcpu)
+{
+    CPURISCVState *env = &cpu->env;
+    uint64_t val;
+    int i, ret;
+
+    for (i = 0; i < ARRAY_SIZE(kvm_multi_ext_cfgs); i++) {
+        KVMCPUConfig *multi_ext_cfg = &kvm_multi_ext_cfgs[i];
+        struct kvm_one_reg reg;
+
+        reg.id = kvm_riscv_reg_id(env, KVM_REG_RISCV_ISA_EXT,
+                                  multi_ext_cfg->kvm_reg_id);
+        reg.addr = (uint64_t)&val;
+        ret = ioctl(kvmcpu->cpufd, KVM_GET_ONE_REG, &reg);
+        if (ret != 0) {
+            if (ret == -EINVAL) {
+                /* Silently default to 'false' if KVM does not support it. */
+                multi_ext_cfg->supported = false;
+                val = false;
+            } else {
+                error_report("Unable to read ISA_EXT KVM register %s, "
+                             "error %d", multi_ext_cfg->name, ret);
+                kvm_riscv_destroy_scratch_vcpu(kvmcpu);
+                exit(EXIT_FAILURE);
+            }
+        } else {
+            multi_ext_cfg->supported = true;
+        }
+
+        kvm_cpu_cfg_set(cpu, multi_ext_cfg, val);
+    }
+}
+
 void kvm_riscv_init_user_properties(Object *cpu_obj)
 {
     RISCVCPU *cpu = RISCV_CPU(cpu_obj);
@@ -542,6 +660,7 @@ void kvm_riscv_init_user_properties(Object *cpu_obj)
     kvm_riscv_add_cpu_user_properties(cpu_obj);
     kvm_riscv_init_machine_ids(cpu, &kvmcpu);
     kvm_riscv_init_misa_ext_mask(cpu, &kvmcpu);
+    kvm_riscv_init_multiext_cfg(cpu, &kvmcpu);
 
     kvm_riscv_destroy_scratch_vcpu(&kvmcpu);
 }
Signature · 1 line
2.41.0
#21 lore raw

PATCH v3 15/19 target/riscv/cpu.c: remove priv_ver check from riscv_isa_string_ext()

In reply to Daniel Henrique Barboza

riscv_isa_string_ext() is being used by riscv_isa_string(), which is then used by boards to retrieve the 'riscv,isa' string to be written in the FDT. All this happens after riscv_cpu_realize(), meaning that we're already past riscv_cpu_validate_set_extensions() and, more important, riscv_cpu_disable_priv_spec_isa_exts().

This means that all extensions that needed to be disabled due to priv_spec mismatch are already disabled. Checking this again during riscv_isa_string_ext() is unneeded. Remove it.

As a bonus, riscv_isa_string_ext() can now be used with the 'host' KVM-only CPU type since it doesn't have a env->priv_ver assigned and it would fail this check for no good reason.

---
 target/riscv/cpu.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
target/riscv/cpu.c +1 -2
index 5a50cb1d0b..b65db165cc 100644
@@ -2001,8 +2001,7 @@ static void riscv_isa_string_ext(RISCVCPU *cpu, char **isa_str,
     int i;
 
     for (i = 0; i < ARRAY_SIZE(isa_edata_arr); i++) {
-        if (cpu->env.priv_ver >= isa_edata_arr[i].min_version &&
-            isa_ext_is_enabled(cpu, &isa_edata_arr[i])) {
+        if (isa_ext_is_enabled(cpu, &isa_edata_arr[i])) {
             new = g_strconcat(old, "_", isa_edata_arr[i].name, NULL);
             g_free(old);
             old = new;
Signature · 1 line
2.41.0
Reviewed-by on v3 15/19
#22 lore raw

Re: [PATCH v3 15/19] target/riscv/cpu.c: remove priv_ver check from riscv_isa_string_ext()

Quoted 36 lines · Daniel Henrique Barboza

riscv_isa_string_ext() is being used by riscv_isa_string(), which is then used by boards to retrieve the 'riscv,isa' string to be written in the FDT. All this happens after riscv_cpu_realize(), meaning that we're already past riscv_cpu_validate_set_extensions() and, more important, riscv_cpu_disable_priv_spec_isa_exts().

This means that all extensions that needed to be disabled due to priv_spec mismatch are already disabled. Checking this again during riscv_isa_string_ext() is unneeded. Remove it.

As a bonus, riscv_isa_string_ext() can now be used with the 'host' KVM-only CPU type since it doesn't have a env->priv_ver assigned and it would fail this check for no good reason.

---
 target/riscv/cpu.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)
target/riscv/cpu.c +1 -2
index 5a50cb1d0b..b65db165cc 100644
@@ -2001,8 +2001,7 @@ static void riscv_isa_string_ext(RISCVCPU *cpu, char **isa_str,
     int i;
 
     for (i = 0; i < ARRAY_SIZE(isa_edata_arr); i++) {
-        if (cpu->env.priv_ver >= isa_edata_arr[i].min_version &&
-            isa_ext_is_enabled(cpu, &isa_edata_arr[i])) {
+        if (isa_ext_is_enabled(cpu, &isa_edata_arr[i])) {
             new = g_strconcat(old, "_", isa_edata_arr[i].name, NULL);
             g_free(old);
             old = new;
Signature · 1 line
2.41.0

Nice!

#23 lore raw

PATCH v3 16/19 target/riscv/cpu.c: create KVM mock properties

In reply to Daniel Henrique Barboza

KVM-specific properties are being created inside target/riscv/kvm.c. But at this moment we're gathering all the remaining properties from TCG and adding them as is when running KVM. This creates a situation where non-KVM properties are setting flags to 'true' due to its default settings (e.g. Zawrs). Users can also freely enable them via command line.

This doesn't impact runtime per se because KVM doesn't care about these flags, but code such as riscv_isa_string_ext() take those flags into account. The result is that, for a KVM guest, setting non-KVM properties will make them appear in the riscv,isa DT.

We want to keep the same API for both TCG and KVM and at the same time, when running KVM, forbid non-KVM extensions to be enabled internally. We accomplish both by changing riscv_cpu_add_user_properties() to add a mock/no-op boolean property for every non-KVM extension in riscv_cpu_extensions[]. Then, when running KVM, users are still free to set extensions at will, we'll treat non-KVM extensions as a no-op, and riscv_isa_string_ext() will not report bogus extensions in the DT.

---
 target/riscv/cpu.c | 36 +++++++++++++++++++++++++++++++++---
 1 file changed, 33 insertions(+), 3 deletions(-)
target/riscv/cpu.c +33 -3
index b65db165cc..f5209f0789 100644
@@ -1720,6 +1720,18 @@ static Property riscv_cpu_extensions[] = {
     DEFINE_PROP_END_OF_LIST(),
 };
 
+
+static void cpu_set_cfg_noop(Object *obj, Visitor *v,
+                             const char *name,
+                             void *opaque, Error **errp)
+{
+    bool value;
+
+    if (!visit_type_bool(v, name, &value, errp)) {
+        return;
+    }
+}
+
 /*
  * Add CPU properties with user-facing flags.
  *
@@ -1738,9 +1750,27 @@ static void riscv_cpu_add_user_properties(Object *obj)
     riscv_cpu_add_misa_properties(obj);
 
     for (prop = riscv_cpu_extensions; prop && prop->name; prop++) {
-        /* Check if KVM didn't create the property already */
-        if (object_property_find(obj, prop->name)) {
-            continue;
+        if (riscv_running_kvm()) {
+            /* Check if KVM didn't create the property already */
+            if (object_property_find(obj, prop->name)) {
+                continue;
+            }
+
+            /*
+             * Set every multi-letter extension that KVM doesn't
+             * know as a no-op. This will allow users to set values
+             * to them while keeping their internal state to 'false'.
+             *
+             * We're giving a pass for non-bool properties since they're
+             * not related to the availability of extensions and can be
+             * safely ignored as is.
+             */
+            if (prop->info == &qdev_prop_bool) {
+                object_property_add(obj, prop->name, "bool",
+                                    NULL, cpu_set_cfg_noop,
+                                    NULL, NULL);
+                continue;
+            }
         }
 
         qdev_property_add_static(dev, prop);
Signature · 1 line
2.41.0
#24 lore raw

Re: [PATCH v3 16/19] target/riscv/cpu.c: create KVM mock properties

Quoted 82 lines · Daniel Henrique Barboza

KVM-specific properties are being created inside target/riscv/kvm.c. But at this moment we're gathering all the remaining properties from TCG and adding them as is when running KVM. This creates a situation where non-KVM properties are setting flags to 'true' due to its default settings (e.g. Zawrs). Users can also freely enable them via command line.

This doesn't impact runtime per se because KVM doesn't care about these flags, but code such as riscv_isa_string_ext() take those flags into account. The result is that, for a KVM guest, setting non-KVM properties will make them appear in the riscv,isa DT.

We want to keep the same API for both TCG and KVM and at the same time, when running KVM, forbid non-KVM extensions to be enabled internally. We accomplish both by changing riscv_cpu_add_user_properties() to add a mock/no-op boolean property for every non-KVM extension in riscv_cpu_extensions[]. Then, when running KVM, users are still free to set extensions at will, we'll treat non-KVM extensions as a no-op, and riscv_isa_string_ext() will not report bogus extensions in the DT.

---
 target/riscv/cpu.c | 36 +++++++++++++++++++++++++++++++++---
 1 file changed, 33 insertions(+), 3 deletions(-)
target/riscv/cpu.c +33 -3
index b65db165cc..f5209f0789 100644
@@ -1720,6 +1720,18 @@ static Property riscv_cpu_extensions[] = {
     DEFINE_PROP_END_OF_LIST(),
 };
 
+
+static void cpu_set_cfg_noop(Object *obj, Visitor *v,
+                             const char *name,
+                             void *opaque, Error **errp)
+{
+    bool value;
+
+    if (!visit_type_bool(v, name, &value, errp)) {
+        return;
+    }
+}
+
 /*
  * Add CPU properties with user-facing flags.
  *
@@ -1738,9 +1750,27 @@ static void riscv_cpu_add_user_properties(Object *obj)
     riscv_cpu_add_misa_properties(obj);
 
     for (prop = riscv_cpu_extensions; prop && prop->name; prop++) {
-        /* Check if KVM didn't create the property already */
-        if (object_property_find(obj, prop->name)) {
-            continue;
+        if (riscv_running_kvm()) {
+            /* Check if KVM didn't create the property already */
+            if (object_property_find(obj, prop->name)) {
+                continue;
+            }
+
+            /*
+             * Set every multi-letter extension that KVM doesn't
+             * know as a no-op. This will allow users to set values
+             * to them while keeping their internal state to 'false'.
+             *
+             * We're giving a pass for non-bool properties since they're
+             * not related to the availability of extensions and can be
+             * safely ignored as is.
+             */
+            if (prop->info == &qdev_prop_bool) {
+                object_property_add(obj, prop->name, "bool",
+                                    NULL, cpu_set_cfg_noop,
+                                    NULL, NULL);
+                continue;
+            }
         }
 
         qdev_property_add_static(dev, prop);
Signature · 1 line
2.41.0

I think we should actually fail with an error when the user tries to enable an extension KVM doesn't support. Otherwise a user may be confused as to why their Zawrs=on didn't provide them a machine with Zawrs. And, when KVM learns how to provide that support to guests (Zawrs is actually on my TODO...), then migrating the same VM to later KVM/QEMU will actually enable the feature, possibly confusing the guest.

So, we should probably just not add any extension properties to KVM guests which can't be enabled. Then, as we add support to KVM, we'll add the properties too.

Thanks,
drew

#25 lore raw

Re: [PATCH v3 16/19] target/riscv/cpu.c: create KVM mock properties

Quoted 92 lines · Andrew Jones
Quoted 82 lines · Daniel Henrique Barboza

KVM-specific properties are being created inside target/riscv/kvm.c. But at this moment we're gathering all the remaining properties from TCG and adding them as is when running KVM. This creates a situation where non-KVM properties are setting flags to 'true' due to its default settings (e.g. Zawrs). Users can also freely enable them via command line.

This doesn't impact runtime per se because KVM doesn't care about these flags, but code such as riscv_isa_string_ext() take those flags into account. The result is that, for a KVM guest, setting non-KVM properties will make them appear in the riscv,isa DT.

We want to keep the same API for both TCG and KVM and at the same time, when running KVM, forbid non-KVM extensions to be enabled internally. We accomplish both by changing riscv_cpu_add_user_properties() to add a mock/no-op boolean property for every non-KVM extension in riscv_cpu_extensions[]. Then, when running KVM, users are still free to set extensions at will, we'll treat non-KVM extensions as a no-op, and riscv_isa_string_ext() will not report bogus extensions in the DT.

---
  target/riscv/cpu.c | 36 +++++++++++++++++++++++++++++++++---
  1 file changed, 33 insertions(+), 3 deletions(-)
target/riscv/cpu.c +33 -3
index b65db165cc..f5209f0789 100644
@@ -1720,6 +1720,18 @@ static Property riscv_cpu_extensions[] = {
      DEFINE_PROP_END_OF_LIST(),
  };
 
+
+static void cpu_set_cfg_noop(Object *obj, Visitor *v,
+                             const char *name,
+                             void *opaque, Error **errp)
+{
+    bool value;
+
+    if (!visit_type_bool(v, name, &value, errp)) {
+        return;
+    }
+}
+
  /*
   * Add CPU properties with user-facing flags.
   *
@@ -1738,9 +1750,27 @@ static void riscv_cpu_add_user_properties(Object *obj)
      riscv_cpu_add_misa_properties(obj);
 
      for (prop = riscv_cpu_extensions; prop && prop->name; prop++) {
-        /* Check if KVM didn't create the property already */
-        if (object_property_find(obj, prop->name)) {
-            continue;
+        if (riscv_running_kvm()) {
+            /* Check if KVM didn't create the property already */
+            if (object_property_find(obj, prop->name)) {
+                continue;
+            }
+
+            /*
+             * Set every multi-letter extension that KVM doesn't
+             * know as a no-op. This will allow users to set values
+             * to them while keeping their internal state to 'false'.
+             *
+             * We're giving a pass for non-bool properties since they're
+             * not related to the availability of extensions and can be
+             * safely ignored as is.
+             */
+            if (prop->info == &qdev_prop_bool) {
+                object_property_add(obj, prop->name, "bool",
+                                    NULL, cpu_set_cfg_noop,
+                                    NULL, NULL);
+                continue;
+            }
          }
 
          qdev_property_add_static(dev, prop);
Signature · 1 line
2.41.0

I think we should actually fail with an error when the user tries to enable an extension KVM doesn't support. Otherwise a user may be confused as to why their Zawrs=on didn't provide them a machine with Zawrs. And, when KVM learns how to provide that support to guests (Zawrs is actually on my TODO...), then migrating the same VM to later KVM/QEMU will actually enable the feature, possibly confusing the guest.

Quoted 3 lines

So, we should probably just not add any extension properties to KVM guests which can't be enabled. Then, as we add support to KVM, we'll add the properties too.

By 'extension properties' do you mean just the flags that enable/disable them, like '-cpu, rawrs=<bool>', or also the other properties related to extensions that KVM might not support, like 'vlen' and 'elen' from RVV? I'd say that it's ok to leave things such as 'vlen' because the user won't be able to enable RVV in KVM anyways.

And what error do we want to throw? With this patch it's easy to just add an Extension zawrs is not available using KVM" error message. Otherwise we can not add the property at all and then QEMU will fail with a "property cpu.X not found" type of error. Both will error out, question is whether we want to be more informative about it.

Thanks,

Daniel

Quoted 3 lines

Thanks,
drew

#26 lore raw

Re: [PATCH v3 16/19] target/riscv/cpu.c: create KVM mock properties

On Fri, Jun 23, 2023 at 11:28:03AM -0300, Daniel Henrique Barboza wrote:
...

Quoted 12 lines
Quoted 11 lines

I think we should actually fail with an error when the user tries to enable an extension KVM doesn't support. Otherwise a user may be confused as to why their Zawrs=on didn't provide them a machine with Zawrs. And, when KVM learns how to provide that support to guests (Zawrs is actually on my TODO...), then migrating the same VM to later KVM/QEMU will actually enable the feature, possibly confusing the guest.

So, we should probably just not add any extension properties to KVM guests which can't be enabled. Then, as we add support to KVM, we'll add the properties too.

Quoted 5 lines

By 'extension properties' do you mean just the flags that enable/disable them, like '-cpu, rawrs=<bool>', or also the other properties related to extensions that KVM might not support, like 'vlen' and 'elen' from RVV? I'd say that it's ok to leave things such as 'vlen' because the user won't be able to enable RVV in KVM anyways.

Properties like 'vlen', which have a dependency on an extension, should probably have their own error checking at cpu finalize features time. I.e. if 'vlen' is present, but not V, then QEMU should complain. I see we don't currently do that, though.

Quoted 6 lines

And what error do we want to throw? With this patch it's easy to just add an Extension zawrs is not available using KVM" error message. Otherwise we can not add the property at all and then QEMU will fail with a "property cpu.X not found" type of error. Both will error out, question is whether we want to be more informative about it.

It's probably best to do the "not available with KVM" error by changing this patch from adding a no-op setter to an error-out setter. That way, our story that a riscv machine is the same for both KVM and TCG still remains, i.e. all properties are still present, but we add the honesty to the story that not everything works with KVM.

Thanks,
drew

#27 lore raw

Re: [PATCH v3 16/19] target/riscv/cpu.c: create KVM mock properties

Quoted 20 lines · Andrew Jones

On Fri, Jun 23, 2023 at 11:28:03AM -0300, Daniel Henrique Barboza wrote:
...

Quoted 12 lines
Quoted 11 lines

I think we should actually fail with an error when the user tries to enable an extension KVM doesn't support. Otherwise a user may be confused as to why their Zawrs=on didn't provide them a machine with Zawrs. And, when KVM learns how to provide that support to guests (Zawrs is actually on my TODO...), then migrating the same VM to later KVM/QEMU will actually enable the feature, possibly confusing the guest.

So, we should probably just not add any extension properties to KVM guests which can't be enabled. Then, as we add support to KVM, we'll add the properties too.

Quoted 5 lines

By 'extension properties' do you mean just the flags that enable/disable them, like '-cpu, rawrs=<bool>', or also the other properties related to extensions that KVM might not support, like 'vlen' and 'elen' from RVV? I'd say that it's ok to leave things such as 'vlen' because the user won't be able to enable RVV in KVM anyways.

Quoted 4 lines

Properties like 'vlen', which have a dependency on an extension, should probably have their own error checking at cpu finalize features time. I.e. if 'vlen' is present, but not V, then QEMU should complain. I see we don't currently do that, though.

We do some checks during realize() in riscv_validate_set_extensions(), via riscv_cpu_validate_v(). However, 'is the user set vlen but not V we should error out' is not possible because we don't have a 'user_set' flag for this option, and we can't say if vlen = 128 is the default, untouched value, or if the user set vlen = 128 in the command line.

I think we're getting closer to a point where we can "export" the KVM config options model to TCG. Let's see if we can pull something off for the next release.

Quoted 8 lines
Quoted 7 lines

And what error do we want to throw? With this patch it's easy to just add an Extension zawrs is not available using KVM" error message. Otherwise we can not add the property at all and then QEMU will fail with a "property cpu.X not found" type of error. Both will error out, question is whether we want to be more informative about it.

Quoted 5 lines

It's probably best to do the "not available with KVM" error by changing this patch from adding a no-op setter to an error-out setter. That way, our story that a riscv machine is the same for both KVM and TCG still remains, i.e. all properties are still present, but we add the honesty to the story that not everything works with KVM.

Got it. I'll add an error message in v4. Thanks,

Daniel

Quoted 3 lines

Thanks,
drew

#28 lore raw

PATCH v3 17/19 target/riscv: update multi-letter extension KVM properties

In reply to Daniel Henrique Barboza

We're now ready to update the multi-letter extensions status for KVM.

kvm_riscv_update_cpu_cfg_isa_ext() is called called during vcpu creation time to verify which user options changes host defaults (via the 'user_set' flag) and tries to write them back to KVM.

Failure to commit a change to KVM is only ignored in case KVM doesn't know about the extension (-EINVAL error code) and the user wanted to disable the given extension. Otherwise we're going to abort the boot process.

---
 target/riscv/kvm.c | 27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)
target/riscv/kvm.c +27 -0
index 376bbc68b4..6eebeb123a 100644
@@ -273,6 +273,32 @@ static void kvm_cpu_set_multi_ext_cfg(Object *obj, Visitor *v,
     kvm_cpu_cfg_set(cpu, multi_ext_cfg, value);
 }
 
+static void kvm_riscv_update_cpu_cfg_isa_ext(RISCVCPU *cpu, CPUState *cs)
+{
+    CPURISCVState *env = &cpu->env;
+    uint64_t id, reg;
+    int i, ret;
+
+    for (i = 0; i < ARRAY_SIZE(kvm_multi_ext_cfgs); i++) {
+        KVMCPUConfig *multi_ext_cfg = &kvm_multi_ext_cfgs[i];
+
+        if (!multi_ext_cfg->user_set) {
+            continue;
+        }
+
+        id = kvm_riscv_reg_id(env, KVM_REG_RISCV_ISA_EXT,
+                              multi_ext_cfg->kvm_reg_id);
+        reg = kvm_cpu_cfg_get(cpu, multi_ext_cfg);
+        ret = kvm_set_one_reg(cs, id, &reg);
+        if (ret != 0) {
+            error_report("Unable to %s extension %s in KVM, error %d",
+                         reg ? "enable" : "disable",
+                         multi_ext_cfg->name, ret);
+            exit(EXIT_FAILURE);
+        }
+    }
+}
+
 static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
 {
     int i;
@@ -792,6 +818,7 @@ int kvm_arch_init_vcpu(CPUState *cs)
     }
 
     kvm_riscv_update_cpu_misa_ext(cpu, cs);
+    kvm_riscv_update_cpu_cfg_isa_ext(cpu, cs);
 
     return ret;
 }
Signature · 1 line
2.41.0
#29 lore raw

PATCH v3 18/19 target/riscv/kvm.c: add kvmconfig_get_cfg_addr() helper

In reply to Daniel Henrique Barboza

There are 2 places in which we need to get a pointer to a certain property of the cpu->cfg struct based on property offset. Next patch will add a couple more.

Create a helper to avoid repeating this code over and over.

---
 target/riscv/kvm.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
target/riscv/kvm.c +7 -4
index 6eebeb123a..e8f6683030 100644
@@ -215,11 +215,15 @@ static KVMCPUConfig kvm_multi_ext_cfgs[] = {
     KVM_EXT_CFG("svpbmt", ext_svpbmt, KVM_RISCV_ISA_EXT_SVPBMT),
 };
 
+static void *kvmconfig_get_cfg_addr(RISCVCPU *cpu, KVMCPUConfig *kvmcfg)
+{
+    return (void *)&cpu->cfg + kvmcfg->offset;
+}
+
 static void kvm_cpu_cfg_set(RISCVCPU *cpu, KVMCPUConfig *multi_ext,
                             uint32_t val)
 {
-    int cpu_cfg_offset = multi_ext->offset;
-    bool *ext_enabled = (void *)&cpu->cfg + cpu_cfg_offset;
+    bool *ext_enabled = kvmconfig_get_cfg_addr(cpu, multi_ext);
 
     *ext_enabled = val;
 }
@@ -227,8 +231,7 @@ static void kvm_cpu_cfg_set(RISCVCPU *cpu, KVMCPUConfig *multi_ext,
 static uint32_t kvm_cpu_cfg_get(RISCVCPU *cpu,
                                 KVMCPUConfig *multi_ext)
 {
-    int cpu_cfg_offset = multi_ext->offset;
-    bool *ext_enabled = (void *)&cpu->cfg + cpu_cfg_offset;
+    bool *ext_enabled = kvmconfig_get_cfg_addr(cpu, multi_ext);
 
     return *ext_enabled;
 }
Signature · 1 line
2.41.0
#30 lore raw

PATCH v3 19/19 target/riscv/kvm.c: read/write (cbom|cboz)_blocksize in KVM

In reply to Daniel Henrique Barboza

If we don't set a proper cbom_blocksize|cboz_blocksize in the FDT the Linux Kernel will fail to detect the availability of the CBOM/CBOZ extensions, regardless of the contents of the 'riscv,isa' DT prop.

The FDT is being written using the cpu->cfg.cbom|z_blocksize attributes, so let's expose them as user properties like it is already done with TCG.

This will also require us to determine proper blocksize values during init() time since the FDT is already created during realize(). We'll take a ride in kvm_riscv_init_multiext_cfg() to do it. Note that we don't need to fetch both cbom and cboz blocksizes every time: check for their parent extensions (icbom and icboz) and only read the blocksizes if needed.

In contrast with cbom/z_blocksize properties from TCG, the user is not able to set any value that is different from the 'host' value when running KVM. KVM can be particularly harsh dealing with it: a ENOTSUPP can be thrown for the mere attempt of executing kvm_set_one_reg() for these 2 regs.

Hopefully, we don't need to call kvm_set_one_reg() for these regs. We'll check if the user input matches the host value in kvm_cpu_set_cbomz_blksize(), the set() accessor for both blocksize properties. We'll fail fast since it's already known to not be supported.

---
 target/riscv/kvm.c | 70 ++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 70 insertions(+)
target/riscv/kvm.c +70 -0
index e8f6683030..58ce42b322 100644
@@ -276,6 +276,42 @@ static void kvm_cpu_set_multi_ext_cfg(Object *obj, Visitor *v,
     kvm_cpu_cfg_set(cpu, multi_ext_cfg, value);
 }
 
+static KVMCPUConfig kvm_cbom_blocksize = {
+    .name = "cbom_blocksize",
+    .offset = CPUCFG(cbom_blocksize),
+    .kvm_reg_id = KVM_REG_RISCV_CONFIG_REG(zicbom_block_size)
+};
+
+static KVMCPUConfig kvm_cboz_blocksize = {
+    .name = "cboz_blocksize",
+    .offset = CPUCFG(cboz_blocksize),
+    .kvm_reg_id = KVM_REG_RISCV_CONFIG_REG(zicboz_block_size)
+};
+
+static void kvm_cpu_set_cbomz_blksize(Object *obj, Visitor *v,
+                                      const char *name,
+                                      void *opaque, Error **errp)
+{
+    KVMCPUConfig *cbomz_cfg = opaque;
+    RISCVCPU *cpu = RISCV_CPU(obj);
+    uint16_t value, *host_val;
+
+    if (!visit_type_uint16(v, name, &value, errp)) {
+        return;
+    }
+
+    host_val = kvmconfig_get_cfg_addr(cpu, cbomz_cfg);
+
+    if (value != *host_val) {
+        error_report("Unable to set %s to a different value than "
+                     "the host (%u)",
+                     cbomz_cfg->name, *host_val);
+        exit(EXIT_FAILURE);
+    }
+
+    cbomz_cfg->user_set = true;
+}
+
 static void kvm_riscv_update_cpu_cfg_isa_ext(RISCVCPU *cpu, CPUState *cs)
 {
     CPURISCVState *env = &cpu->env;
@@ -329,6 +365,14 @@ static void kvm_riscv_add_cpu_user_properties(Object *cpu_obj)
                             kvm_cpu_set_multi_ext_cfg,
                             NULL, multi_cfg);
     }
+
+    object_property_add(cpu_obj, "cbom_blocksize", "uint16",
+                        NULL, kvm_cpu_set_cbomz_blksize,
+                        NULL, &kvm_cbom_blocksize);
+
+    object_property_add(cpu_obj, "cboz_blocksize", "uint16",
+                        NULL, kvm_cpu_set_cbomz_blksize,
+                        NULL, &kvm_cboz_blocksize);
 }
 
 static int kvm_riscv_get_regs_core(CPUState *cs)
@@ -644,6 +688,24 @@ static void kvm_riscv_init_misa_ext_mask(RISCVCPU *cpu,
     env->misa_ext = env->misa_ext_mask;
 }
 
+static void kvm_riscv_read_cbomz_blksize(RISCVCPU *cpu, KVMScratchCPU *kvmcpu,
+                                         KVMCPUConfig *cbomz_cfg)
+{
+    CPURISCVState *env = &cpu->env;
+    struct kvm_one_reg reg;
+    int ret;
+
+    reg.id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
+                              cbomz_cfg->kvm_reg_id);
+    reg.addr = (uint64_t)kvmconfig_get_cfg_addr(cpu, cbomz_cfg);
+    ret = ioctl(kvmcpu->cpufd, KVM_GET_ONE_REG, &reg);
+    if (ret != 0) {
+        error_report("Unable to read KVM reg %s, error %d",
+                     cbomz_cfg->name, ret);
+        exit(EXIT_FAILURE);
+    }
+}
+
 static void kvm_riscv_init_multiext_cfg(RISCVCPU *cpu, KVMScratchCPU *kvmcpu)
 {
     CPURISCVState *env = &cpu->env;
@@ -675,6 +737,14 @@ static void kvm_riscv_init_multiext_cfg(RISCVCPU *cpu, KVMScratchCPU *kvmcpu)
 
         kvm_cpu_cfg_set(cpu, multi_ext_cfg, val);
     }
+
+    if (cpu->cfg.ext_icbom) {
+        kvm_riscv_read_cbomz_blksize(cpu, kvmcpu, &kvm_cbom_blocksize);
+    }
+
+    if (cpu->cfg.ext_icboz) {
+        kvm_riscv_read_cbomz_blksize(cpu, kvmcpu, &kvm_cboz_blocksize);
+    }
 }
 
 void kvm_riscv_init_user_properties(Object *cpu_obj)
Signature · 1 line
2.41.0