[Qemu-devel] [PATCH v4 12/12] spapr kvm vfio: enable in-kernel acceleration

[Alexey Kardashevskiy]

This enables in-kernel support for DMA operations on VFIO PHBs.
This creates a "SPAPR TCE IOMMU" KVM device and initializes it
with LIOBN and IOMMU fd. Once enabled, it keeps H_PUT_TCE,
H_PUT_TCE_INDIRECT, H_STUFF_TCE hypercalls in the kernel so
their QEMU handlers won't be called.

Signed-off-by: Alexey Kardashevskiy <address@hidden>
---
 hw/ppc/spapr_iommu.c |  4 ++++
 target-ppc/kvm.c     | 47 +++++++++++++++++++++++++++++++++++++++++++++++
 target-ppc/kvm_ppc.h | 13 +++++++++++++
 3 files changed, 64 insertions(+)

diff --git a/hw/ppc/spapr_iommu.c b/hw/ppc/spapr_iommu.c
index d91dee5..32255fc 100644
--- a/hw/ppc/spapr_iommu.c
+++ b/hw/ppc/spapr_iommu.c
@@ -258,6 +258,7 @@ sPAPRTCETable *spapr_vfio_new_table(DeviceState *owner, 
uint32_t liobn,
                                     int group_fd)
 {
     sPAPRTCETable *tcet;
+    int fd;
 
     if (spapr_tce_find_by_liobn(liobn)) {
         fprintf(stderr, "Attempted to create TCE table with duplicate"
@@ -265,8 +266,11 @@ sPAPRTCETable *spapr_vfio_new_table(DeviceState *owner, 
uint32_t liobn,
         return NULL;
     }
 
+    fd = kvmppc_create_spapr_tce_iommu(liobn, group_fd);
+
     tcet = SPAPR_TCE_TABLE(object_new(TYPE_SPAPR_TCE_TABLE_VFIO));
     tcet->liobn = liobn;
+    tcet->fd = fd;
     object_property_add_child(OBJECT(owner), "tce-table", OBJECT(tcet), NULL);
 
     object_property_set_bool(OBJECT(tcet), true, "realized", NULL);
diff --git a/target-ppc/kvm.c b/target-ppc/kvm.c
index 5e72c22..ec7556f 100644
--- a/target-ppc/kvm.c
+++ b/target-ppc/kvm.c
@@ -60,6 +60,7 @@ static int cap_booke_sregs;
 static int cap_ppc_smt;
 static int cap_ppc_rma;
 static int cap_spapr_tce;
+static int cap_spapr_tce_iommu;
 static int cap_hior;
 static int cap_one_reg;
 static int cap_epr;
@@ -96,6 +97,7 @@ int kvm_arch_init(KVMState *s)
     cap_ppc_smt = kvm_check_extension(s, KVM_CAP_PPC_SMT);
     cap_ppc_rma = kvm_check_extension(s, KVM_CAP_PPC_RMA);
     cap_spapr_tce = kvm_check_extension(s, KVM_CAP_SPAPR_TCE);
+    cap_spapr_tce_iommu = kvm_check_extension(s, KVM_CAP_SPAPR_TCE_IOMMU);
     cap_one_reg = kvm_check_extension(s, KVM_CAP_ONE_REG);
     cap_hior = kvm_check_extension(s, KVM_CAP_PPC_HIOR);
     cap_epr = kvm_check_extension(s, KVM_CAP_PPC_EPR);
@@ -1662,6 +1664,51 @@ int kvmppc_remove_spapr_tce(void *table, int fd, 
uint32_t window_size)
     return 0;
 }
 
+int kvmppc_create_spapr_tce_iommu(uint32_t liobn, int group_fd)
+{
+    int rc;
+    struct kvm_create_spapr_tce_iommu_linkage args = {
+        .liobn = liobn,
+        .fd = group_fd
+    };
+    struct kvm_device_attr attr = {
+        .flags = 0,
+        .group = KVM_DEV_SPAPR_TCE_IOMMU_ATTR_LINKAGE,
+        .addr = (uint64_t)(uintptr_t)&args,
+    };
+    struct kvm_create_device kcd = {
+        .type = KVM_DEV_TYPE_SPAPR_TCE_IOMMU,
+        .flags = 0,
+    };
+
+    if (!kvm_enabled() || !cap_spapr_tce_iommu) {
+        fprintf(stderr, "KVM VFIO: TCE IOMMU capability is not present, DMA 
may be slow\n");
+        return -1;
+    }
+
+    rc = kvm_vm_ioctl(kvm_state, KVM_CREATE_DEVICE, &kcd);
+    if (rc < 0) {
+        fprintf(stderr, "Error on KVM_CREATE_DEVICE for SPAPR TCE IOMMU\n");
+        return rc;
+    }
+    rc = ioctl(kcd.fd, KVM_SET_DEVICE_ATTR, &attr);
+    if (rc < 0) {
+        fprintf(stderr, "KVM VFIO: Failed to create TCE table for liobn 0x%x, 
ret = %d, DMA may be slow\n",
+                liobn, rc);
+    }
+
+    return kcd.fd;
+}
+
+int kvmppc_remove_spapr_tce_iommu(int fd)
+{
+    if (fd < 0) {
+        return -1;
+    }
+
+    return close(fd);
+}
+
 int kvmppc_reset_htab(int shift_hint)
 {
     uint32_t shift = shift_hint;
diff --git a/target-ppc/kvm_ppc.h b/target-ppc/kvm_ppc.h
index 5f78e4b..35d3325 100644
--- a/target-ppc/kvm_ppc.h
+++ b/target-ppc/kvm_ppc.h
@@ -33,6 +33,8 @@ int kvmppc_booke_watchdog_enable(PowerPCCPU *cpu);
 off_t kvmppc_alloc_rma(const char *name, MemoryRegion *sysmem);
 void *kvmppc_create_spapr_tce(uint32_t liobn, uint32_t window_size, int *pfd);
 int kvmppc_remove_spapr_tce(void *table, int pfd, uint32_t window_size);
+int kvmppc_create_spapr_tce_iommu(uint32_t liobn, int group_fd);
+int kvmppc_remove_spapr_tce_iommu(int fd);
 int kvmppc_reset_htab(int shift_hint);
 uint64_t kvmppc_rma_size(uint64_t current_size, unsigned int hash_shift);
 #endif /* !CONFIG_USER_ONLY */
@@ -137,6 +139,17 @@ static inline int kvmppc_remove_spapr_tce(void *table, int 
pfd,
     return -1;
 }
 
+static inline int kvmppc_create_spapr_tce_iommu(uint32_t liobn,
+                                                uint32_t iommu_id)
+{
+    return -1;
+}
+
+static inline int kvmppc_remove_spapr_tce_iommu(int fd)
+{
+    return -1;
+}
+
 static inline int kvmppc_reset_htab(int shift_hint)
 {
     return -1;
-- 
1.8.4.rc4