KVM: PPC: Book3S HV: XIVE: Add controls for the EQ configuration
authorCédric Le Goater <clg@kaod.org>
Thu, 18 Apr 2019 10:39:31 +0000 (12:39 +0200)
committerPaul Mackerras <paulus@ozlabs.org>
Tue, 30 Apr 2019 09:35:16 +0000 (19:35 +1000)
These controls will be used by the H_INT_SET_QUEUE_CONFIG and
H_INT_GET_QUEUE_CONFIG hcalls from QEMU to configure the underlying
Event Queue in the XIVE IC. They will also be used to restore the
configuration of the XIVE EQs and to capture the internal run-time
state of the EQs. Both 'get' and 'set' rely on an OPAL call to access
the EQ toggle bit and EQ index which are updated by the XIVE IC when
event notifications are enqueued in the EQ.

The value of the guest physical address of the event queue is saved in
the XIVE internal xive_q structure for later use. That is when
migration needs to mark the EQ pages dirty to capture a consistent
memory state of the VM.

To be noted that H_INT_SET_QUEUE_CONFIG does not require the extra
OPAL call setting the EQ toggle bit and EQ index to configure the EQ,
but restoring the EQ state will.

Signed-off-by: Cédric Le Goater <clg@kaod.org>
Reviewed-by: David Gibson <david@gibson.dropbear.id.au>
Signed-off-by: Paul Mackerras <paulus@ozlabs.org>
Documentation/virtual/kvm/devices/xive.txt
arch/powerpc/include/asm/xive.h
arch/powerpc/include/uapi/asm/kvm.h
arch/powerpc/kvm/book3s_xive.c
arch/powerpc/kvm/book3s_xive.h
arch/powerpc/kvm/book3s_xive_native.c

index 33c64b2cdbe8534761f10aeff22149e9022838b5..cc13bfd5cf53f43fb7e6864940b3f60af3ddee2f 100644 (file)
@@ -53,3 +53,37 @@ the legacy interrupt mode, referred as XICS (POWER7/8).
     -ENXIO:  CPU event queues not configured or configuration of the
              underlying HW interrupt failed
     -EBUSY:  No CPU available to serve interrupt
+
+  4. KVM_DEV_XIVE_GRP_EQ_CONFIG (read-write)
+  Configures an event queue of a CPU
+  Attributes:
+    EQ descriptor identifier (64-bit)
+  The EQ descriptor identifier is a tuple (server, priority) :
+  bits:     | 63   ....  32 | 31 .. 3 |  2 .. 0
+  values:   |    unused     |  server | priority
+  The kvm_device_attr.addr points to :
+    struct kvm_ppc_xive_eq {
+       __u32 flags;
+       __u32 qshift;
+       __u64 qaddr;
+       __u32 qtoggle;
+       __u32 qindex;
+       __u8  pad[40];
+    };
+  - flags: queue flags
+    KVM_XIVE_EQ_ALWAYS_NOTIFY (required)
+       forces notification without using the coalescing mechanism
+       provided by the XIVE END ESBs.
+  - qshift: queue size (power of 2)
+  - qaddr: real address of queue
+  - qtoggle: current queue toggle bit
+  - qindex: current queue index
+  - pad: reserved for future use
+  Errors:
+    -ENOENT: Invalid CPU number
+    -EINVAL: Invalid priority
+    -EINVAL: Invalid flags
+    -EINVAL: Invalid queue size
+    -EINVAL: Invalid queue address
+    -EFAULT: Invalid user pointer for attr->addr.
+    -EIO:    Configuration of the underlying HW failed
index b579a943407bbc0a63c34d53fc73662e9e99db7a..c4e88abd3b67e915f3fa4db602f76324cb44e4bd 100644 (file)
@@ -73,6 +73,8 @@ struct xive_q {
        u32                     esc_irq;
        atomic_t                count;
        atomic_t                pending_count;
+       u64                     guest_qaddr;
+       u32                     guest_qshift;
 };
 
 /* Global enable flags for the XIVE support */
index e8161e21629b1cd7e428e767ccef765ffe11a396..85005400fd86c88847280cf8e2e0036af9149961 100644 (file)
@@ -681,6 +681,7 @@ struct kvm_ppc_cpu_char {
 #define KVM_DEV_XIVE_GRP_CTRL          1
 #define KVM_DEV_XIVE_GRP_SOURCE                2       /* 64-bit source identifier */
 #define KVM_DEV_XIVE_GRP_SOURCE_CONFIG 3       /* 64-bit source identifier */
+#define KVM_DEV_XIVE_GRP_EQ_CONFIG     4       /* 64-bit EQ identifier */
 
 /* Layout of 64-bit XIVE source attribute values */
 #define KVM_XIVE_LEVEL_SENSITIVE       (1ULL << 0)
@@ -696,4 +697,22 @@ struct kvm_ppc_cpu_char {
 #define KVM_XIVE_SOURCE_EISN_SHIFT     33
 #define KVM_XIVE_SOURCE_EISN_MASK      0xfffffffe00000000ULL
 
+/* Layout of 64-bit EQ identifier */
+#define KVM_XIVE_EQ_PRIORITY_SHIFT     0
+#define KVM_XIVE_EQ_PRIORITY_MASK      0x7
+#define KVM_XIVE_EQ_SERVER_SHIFT       3
+#define KVM_XIVE_EQ_SERVER_MASK                0xfffffff8ULL
+
+/* Layout of EQ configuration values (64 bytes) */
+struct kvm_ppc_xive_eq {
+       __u32 flags;
+       __u32 qshift;
+       __u64 qaddr;
+       __u32 qtoggle;
+       __u32 qindex;
+       __u8  pad[40];
+};
+
+#define KVM_XIVE_EQ_ALWAYS_NOTIFY      0x00000001
+
 #endif /* __LINUX_KVM_POWERPC_H */
index e09f3addffe5aa6f4b21a32952a57ecb29eeeea8..c1b7aa7dbc282558ad199d99e1b5c85279ee1837 100644 (file)
@@ -166,7 +166,8 @@ static irqreturn_t xive_esc_irq(int irq, void *data)
        return IRQ_HANDLED;
 }
 
-static int xive_attach_escalation(struct kvm_vcpu *vcpu, u8 prio)
+int kvmppc_xive_attach_escalation(struct kvm_vcpu *vcpu, u8 prio,
+                                 bool single_escalation)
 {
        struct kvmppc_xive_vcpu *xc = vcpu->arch.xive_vcpu;
        struct xive_q *q = &xc->queues[prio];
@@ -185,7 +186,7 @@ static int xive_attach_escalation(struct kvm_vcpu *vcpu, u8 prio)
                return -EIO;
        }
 
-       if (xc->xive->single_escalation)
+       if (single_escalation)
                name = kasprintf(GFP_KERNEL, "kvm-%d-%d",
                                 vcpu->kvm->arch.lpid, xc->server_num);
        else
@@ -217,7 +218,7 @@ static int xive_attach_escalation(struct kvm_vcpu *vcpu, u8 prio)
         * interrupt, thus leaving it effectively masked after
         * it fires once.
         */
-       if (xc->xive->single_escalation) {
+       if (single_escalation) {
                struct irq_data *d = irq_get_irq_data(xc->esc_virq[prio]);
                struct xive_irq_data *xd = irq_data_get_irq_handler_data(d);
 
@@ -291,7 +292,8 @@ static int xive_check_provisioning(struct kvm *kvm, u8 prio)
                        continue;
                rc = xive_provision_queue(vcpu, prio);
                if (rc == 0 && !xive->single_escalation)
-                       xive_attach_escalation(vcpu, prio);
+                       kvmppc_xive_attach_escalation(vcpu, prio,
+                                                     xive->single_escalation);
                if (rc)
                        return rc;
        }
@@ -1214,7 +1216,8 @@ int kvmppc_xive_connect_vcpu(struct kvm_device *dev,
                if (xive->qmap & (1 << i)) {
                        r = xive_provision_queue(vcpu, i);
                        if (r == 0 && !xive->single_escalation)
-                               xive_attach_escalation(vcpu, i);
+                               kvmppc_xive_attach_escalation(
+                                       vcpu, i, xive->single_escalation);
                        if (r)
                                goto bail;
                } else {
@@ -1229,7 +1232,7 @@ int kvmppc_xive_connect_vcpu(struct kvm_device *dev,
        }
 
        /* If not done above, attach priority 0 escalation */
-       r = xive_attach_escalation(vcpu, 0);
+       r = kvmppc_xive_attach_escalation(vcpu, 0, xive->single_escalation);
        if (r)
                goto bail;
 
index ae26fe653d98952b229bc413028e4c913027ea31..622f594d93e1de20be3888b3cd7ca9a588e2d10a 100644 (file)
@@ -272,6 +272,8 @@ struct kvmppc_xive_src_block *kvmppc_xive_create_src_block(
        struct kvmppc_xive *xive, int irq);
 void kvmppc_xive_free_sources(struct kvmppc_xive_src_block *sb);
 int kvmppc_xive_select_target(struct kvm *kvm, u32 *server, u8 prio);
+int kvmppc_xive_attach_escalation(struct kvm_vcpu *vcpu, u8 prio,
+                                 bool single_escalation);
 
 #endif /* CONFIG_KVM_XICS */
 #endif /* _KVM_PPC_BOOK3S_XICS_H */
index 492825a3595803daaae0662297efcf6bf01605cd..3e7cdcacc932950300f6986f86f2c9fb15baa73e 100644 (file)
@@ -335,6 +335,243 @@ static int kvmppc_xive_native_set_source_config(struct kvmppc_xive *xive,
                                                       priority, masked, eisn);
 }
 
+static int xive_native_validate_queue_size(u32 qshift)
+{
+       /*
+        * We only support 64K pages for the moment. This is also
+        * advertised in the DT property "ibm,xive-eq-sizes"
+        */
+       switch (qshift) {
+       case 0: /* EQ reset */
+       case 16:
+               return 0;
+       case 12:
+       case 21:
+       case 24:
+       default:
+               return -EINVAL;
+       }
+}
+
+static int kvmppc_xive_native_set_queue_config(struct kvmppc_xive *xive,
+                                              long eq_idx, u64 addr)
+{
+       struct kvm *kvm = xive->kvm;
+       struct kvm_vcpu *vcpu;
+       struct kvmppc_xive_vcpu *xc;
+       void __user *ubufp = (void __user *) addr;
+       u32 server;
+       u8 priority;
+       struct kvm_ppc_xive_eq kvm_eq;
+       int rc;
+       __be32 *qaddr = 0;
+       struct page *page;
+       struct xive_q *q;
+       gfn_t gfn;
+       unsigned long page_size;
+
+       /*
+        * Demangle priority/server tuple from the EQ identifier
+        */
+       priority = (eq_idx & KVM_XIVE_EQ_PRIORITY_MASK) >>
+               KVM_XIVE_EQ_PRIORITY_SHIFT;
+       server = (eq_idx & KVM_XIVE_EQ_SERVER_MASK) >>
+               KVM_XIVE_EQ_SERVER_SHIFT;
+
+       if (copy_from_user(&kvm_eq, ubufp, sizeof(kvm_eq)))
+               return -EFAULT;
+
+       vcpu = kvmppc_xive_find_server(kvm, server);
+       if (!vcpu) {
+               pr_err("Can't find server %d\n", server);
+               return -ENOENT;
+       }
+       xc = vcpu->arch.xive_vcpu;
+
+       if (priority != xive_prio_from_guest(priority)) {
+               pr_err("Trying to restore invalid queue %d for VCPU %d\n",
+                      priority, server);
+               return -EINVAL;
+       }
+       q = &xc->queues[priority];
+
+       pr_devel("%s VCPU %d priority %d fl:%x shift:%d addr:%llx g:%d idx:%d\n",
+                __func__, server, priority, kvm_eq.flags,
+                kvm_eq.qshift, kvm_eq.qaddr, kvm_eq.qtoggle, kvm_eq.qindex);
+
+       /*
+        * sPAPR specifies a "Unconditional Notify (n) flag" for the
+        * H_INT_SET_QUEUE_CONFIG hcall which forces notification
+        * without using the coalescing mechanisms provided by the
+        * XIVE END ESBs. This is required on KVM as notification
+        * using the END ESBs is not supported.
+        */
+       if (kvm_eq.flags != KVM_XIVE_EQ_ALWAYS_NOTIFY) {
+               pr_err("invalid flags %d\n", kvm_eq.flags);
+               return -EINVAL;
+       }
+
+       rc = xive_native_validate_queue_size(kvm_eq.qshift);
+       if (rc) {
+               pr_err("invalid queue size %d\n", kvm_eq.qshift);
+               return rc;
+       }
+
+       /* reset queue and disable queueing */
+       if (!kvm_eq.qshift) {
+               q->guest_qaddr  = 0;
+               q->guest_qshift = 0;
+
+               rc = xive_native_configure_queue(xc->vp_id, q, priority,
+                                                NULL, 0, true);
+               if (rc) {
+                       pr_err("Failed to reset queue %d for VCPU %d: %d\n",
+                              priority, xc->server_num, rc);
+                       return rc;
+               }
+
+               if (q->qpage) {
+                       put_page(virt_to_page(q->qpage));
+                       q->qpage = NULL;
+               }
+
+               return 0;
+       }
+
+       if (kvm_eq.qaddr & ((1ull << kvm_eq.qshift) - 1)) {
+               pr_err("queue page is not aligned %llx/%llx\n", kvm_eq.qaddr,
+                      1ull << kvm_eq.qshift);
+               return -EINVAL;
+       }
+
+       gfn = gpa_to_gfn(kvm_eq.qaddr);
+       page = gfn_to_page(kvm, gfn);
+       if (is_error_page(page)) {
+               pr_err("Couldn't get queue page %llx!\n", kvm_eq.qaddr);
+               return -EINVAL;
+       }
+
+       page_size = kvm_host_page_size(kvm, gfn);
+       if (1ull << kvm_eq.qshift > page_size) {
+               pr_warn("Incompatible host page size %lx!\n", page_size);
+               return -EINVAL;
+       }
+
+       qaddr = page_to_virt(page) + (kvm_eq.qaddr & ~PAGE_MASK);
+
+       /*
+        * Backup the queue page guest address to the mark EQ page
+        * dirty for migration.
+        */
+       q->guest_qaddr  = kvm_eq.qaddr;
+       q->guest_qshift = kvm_eq.qshift;
+
+        /*
+         * Unconditional Notification is forced by default at the
+         * OPAL level because the use of END ESBs is not supported by
+         * Linux.
+         */
+       rc = xive_native_configure_queue(xc->vp_id, q, priority,
+                                        (__be32 *) qaddr, kvm_eq.qshift, true);
+       if (rc) {
+               pr_err("Failed to configure queue %d for VCPU %d: %d\n",
+                      priority, xc->server_num, rc);
+               put_page(page);
+               return rc;
+       }
+
+       /*
+        * Only restore the queue state when needed. When doing the
+        * H_INT_SET_SOURCE_CONFIG hcall, it should not.
+        */
+       if (kvm_eq.qtoggle != 1 || kvm_eq.qindex != 0) {
+               rc = xive_native_set_queue_state(xc->vp_id, priority,
+                                                kvm_eq.qtoggle,
+                                                kvm_eq.qindex);
+               if (rc)
+                       goto error;
+       }
+
+       rc = kvmppc_xive_attach_escalation(vcpu, priority,
+                                          xive->single_escalation);
+error:
+       if (rc)
+               kvmppc_xive_native_cleanup_queue(vcpu, priority);
+       return rc;
+}
+
+static int kvmppc_xive_native_get_queue_config(struct kvmppc_xive *xive,
+                                              long eq_idx, u64 addr)
+{
+       struct kvm *kvm = xive->kvm;
+       struct kvm_vcpu *vcpu;
+       struct kvmppc_xive_vcpu *xc;
+       struct xive_q *q;
+       void __user *ubufp = (u64 __user *) addr;
+       u32 server;
+       u8 priority;
+       struct kvm_ppc_xive_eq kvm_eq;
+       u64 qaddr;
+       u64 qshift;
+       u64 qeoi_page;
+       u32 escalate_irq;
+       u64 qflags;
+       int rc;
+
+       /*
+        * Demangle priority/server tuple from the EQ identifier
+        */
+       priority = (eq_idx & KVM_XIVE_EQ_PRIORITY_MASK) >>
+               KVM_XIVE_EQ_PRIORITY_SHIFT;
+       server = (eq_idx & KVM_XIVE_EQ_SERVER_MASK) >>
+               KVM_XIVE_EQ_SERVER_SHIFT;
+
+       vcpu = kvmppc_xive_find_server(kvm, server);
+       if (!vcpu) {
+               pr_err("Can't find server %d\n", server);
+               return -ENOENT;
+       }
+       xc = vcpu->arch.xive_vcpu;
+
+       if (priority != xive_prio_from_guest(priority)) {
+               pr_err("invalid priority for queue %d for VCPU %d\n",
+                      priority, server);
+               return -EINVAL;
+       }
+       q = &xc->queues[priority];
+
+       memset(&kvm_eq, 0, sizeof(kvm_eq));
+
+       if (!q->qpage)
+               return 0;
+
+       rc = xive_native_get_queue_info(xc->vp_id, priority, &qaddr, &qshift,
+                                       &qeoi_page, &escalate_irq, &qflags);
+       if (rc)
+               return rc;
+
+       kvm_eq.flags = 0;
+       if (qflags & OPAL_XIVE_EQ_ALWAYS_NOTIFY)
+               kvm_eq.flags |= KVM_XIVE_EQ_ALWAYS_NOTIFY;
+
+       kvm_eq.qshift = q->guest_qshift;
+       kvm_eq.qaddr  = q->guest_qaddr;
+
+       rc = xive_native_get_queue_state(xc->vp_id, priority, &kvm_eq.qtoggle,
+                                        &kvm_eq.qindex);
+       if (rc)
+               return rc;
+
+       pr_devel("%s VCPU %d priority %d fl:%x shift:%d addr:%llx g:%d idx:%d\n",
+                __func__, server, priority, kvm_eq.flags,
+                kvm_eq.qshift, kvm_eq.qaddr, kvm_eq.qtoggle, kvm_eq.qindex);
+
+       if (copy_to_user(ubufp, &kvm_eq, sizeof(kvm_eq)))
+               return -EFAULT;
+
+       return 0;
+}
+
 static int kvmppc_xive_native_set_attr(struct kvm_device *dev,
                                       struct kvm_device_attr *attr)
 {
@@ -349,6 +586,9 @@ static int kvmppc_xive_native_set_attr(struct kvm_device *dev,
        case KVM_DEV_XIVE_GRP_SOURCE_CONFIG:
                return kvmppc_xive_native_set_source_config(xive, attr->attr,
                                                            attr->addr);
+       case KVM_DEV_XIVE_GRP_EQ_CONFIG:
+               return kvmppc_xive_native_set_queue_config(xive, attr->attr,
+                                                          attr->addr);
        }
        return -ENXIO;
 }
@@ -356,6 +596,13 @@ static int kvmppc_xive_native_set_attr(struct kvm_device *dev,
 static int kvmppc_xive_native_get_attr(struct kvm_device *dev,
                                       struct kvm_device_attr *attr)
 {
+       struct kvmppc_xive *xive = dev->private;
+
+       switch (attr->group) {
+       case KVM_DEV_XIVE_GRP_EQ_CONFIG:
+               return kvmppc_xive_native_get_queue_config(xive, attr->attr,
+                                                          attr->addr);
+       }
        return -ENXIO;
 }
 
@@ -371,6 +618,8 @@ static int kvmppc_xive_native_has_attr(struct kvm_device *dev,
                    attr->attr < KVMPPC_XIVE_NR_IRQS)
                        return 0;
                break;
+       case KVM_DEV_XIVE_GRP_EQ_CONFIG:
+               return 0;
        }
        return -ENXIO;
 }