Add api for communication betwen primary and secondary VMs.
diff --git a/inc/api.h b/inc/api.h
index edcd9e5..851c2da 100644
--- a/inc/api.h
+++ b/inc/api.h
@@ -9,9 +9,18 @@
extern uint32_t secondary_vm_count;
extern struct vm primary_vm;
+struct vcpu *api_switch_to_primary(size_t primary_retval,
+ enum vcpu_state secondary_state);
+
int32_t api_vm_get_count(void);
int32_t api_vcpu_get_count(uint32_t vm_idx);
int32_t api_vcpu_run(uint32_t vm_idx, uint32_t vcpu_idx, struct vcpu **next);
struct vcpu *api_wait_for_interrupt(void);
+int32_t api_vm_configure(paddr_t send, paddr_t recv);
+
+int32_t api_rpc_request(uint32_t vm_idx, size_t size);
+int32_t api_rpc_read_request(bool block, struct vcpu **next);
+int32_t api_rpc_reply(size_t size, bool ack, struct vcpu **next);
+int32_t api_rpc_ack(void);
#endif /* _API_H */
diff --git a/inc/cpu.h b/inc/cpu.h
index 30433ad..c1ab1c6 100644
--- a/inc/cpu.h
+++ b/inc/cpu.h
@@ -8,11 +8,20 @@
#include "arch_cpu.h"
#include "spinlock.h"
+enum vcpu_state {
+ vcpu_state_off,
+ vcpu_state_ready,
+ vcpu_state_running,
+ vcpu_state_blocked_rpc,
+ vcpu_state_blocked_interrupt,
+};
+
struct vcpu {
struct spinlock lock;
- bool is_on;
- struct arch_regs regs;
+ enum vcpu_state state;
struct vm *vm;
+ struct vcpu *rpc_next;
+ struct arch_regs regs;
};
/* TODO: Update alignment such that cpus are in different cache lines. */
diff --git a/inc/mm.h b/inc/mm.h
index 7eb3169..df027dd 100644
--- a/inc/mm.h
+++ b/inc/mm.h
@@ -45,6 +45,7 @@
paddr_t paddr, int mode);
bool mm_ptable_map_page(struct mm_ptable *t, vaddr_t va, paddr_t pa, int mode);
bool mm_ptable_unmap(struct mm_ptable *t, vaddr_t begin, vaddr_t end, int mode);
+bool mm_ptable_is_mapped(struct mm_ptable *t, vaddr_t addr, int mode);
void mm_ptable_defrag(struct mm_ptable *t, int mode);
bool mm_ptable_unmap_hypervisor(struct mm_ptable *t, int mode);
diff --git a/inc/vm.h b/inc/vm.h
index 847d5b7..2fb3bbf 100644
--- a/inc/vm.h
+++ b/inc/vm.h
@@ -4,7 +4,23 @@
#include "cpu.h"
#include "mm.h"
+enum rpc_state {
+ rpc_state_idle,
+ rpc_state_pending,
+ rpc_state_inflight,
+};
+
+struct rpc {
+ enum rpc_state state;
+ int16_t recv_bytes;
+ void *recv;
+ const void *send;
+ struct vcpu *recv_waiter;
+};
+
struct vm {
+ struct spinlock lock;
+ struct rpc rpc;
struct mm_ptable ptable;
uint32_t vcpu_count;
struct vcpu vcpus[MAX_CPUS];
diff --git a/src/api.c b/src/api.c
index 0030473..ae429f0 100644
--- a/src/api.c
+++ b/src/api.c
@@ -1,6 +1,7 @@
#include "api.h"
#include "arch_api.h"
+#include "std.h"
#include "vm.h"
struct vm secondary_vm[MAX_VMS];
@@ -8,6 +9,32 @@
struct vm primary_vm;
/**
+ * Switches the physical CPU back to the corresponding vcpu of the primary VM.
+ */
+struct vcpu *api_switch_to_primary(size_t primary_retval,
+ enum vcpu_state secondary_state)
+{
+ struct vcpu *vcpu = cpu()->current;
+ struct vcpu *next = &primary_vm.vcpus[cpu_index(cpu())];
+
+ /* Switch back to primary VM. */
+ vm_set_current(&primary_vm);
+
+ /*
+ * Inidicate to primary VM that this vcpu blocked waiting for an
+ * interrupt.
+ */
+ arch_regs_set_retval(&next->regs, primary_retval);
+
+ /* Mark the vcpu as waiting. */
+ sl_lock(&vcpu->lock);
+ vcpu->state = secondary_state;
+ sl_unlock(&vcpu->lock);
+
+ return next;
+}
+
+/**
* Returns the number of VMs configured to run.
*/
int32_t api_vm_get_count(void)
@@ -32,8 +59,9 @@
*/
int32_t api_vcpu_run(uint32_t vm_idx, uint32_t vcpu_idx, struct vcpu **next)
{
- struct vm *vm = secondary_vm + vm_idx;
+ struct vm *vm;
struct vcpu *vcpu;
+ int32_t ret;
/* Only the primary VM can switch vcpus. */
if (cpu()->current->vm != &primary_vm) {
@@ -44,15 +72,25 @@
return HF_VCPU_WAIT_FOR_INTERRUPT;
}
- vcpu = vm->vcpus + vcpu_idx;
- if (vcpu_idx >= vm->vcpu_count || !vcpu->is_on) {
+ vm = secondary_vm + vm_idx;
+ if (vcpu_idx >= vm->vcpu_count) {
return HF_VCPU_WAIT_FOR_INTERRUPT;
}
- vm_set_current(vm);
- *next = vcpu;
+ vcpu = vm->vcpus + vcpu_idx;
- return HF_VCPU_YIELD;
+ sl_lock(&vcpu->lock);
+ if (vcpu->state != vcpu_state_ready) {
+ ret = HF_VCPU_WAIT_FOR_INTERRUPT;
+ } else {
+ vcpu->state = vcpu_state_running;
+ vm_set_current(vm);
+ *next = vcpu;
+ ret = HF_VCPU_YIELD;
+ }
+ sl_unlock(&vcpu->lock);
+
+ return ret;
}
/**
@@ -61,16 +99,289 @@
*/
struct vcpu *api_wait_for_interrupt(void)
{
- struct vcpu *vcpu = &primary_vm.vcpus[cpu_index(cpu())];
+ return api_switch_to_primary(HF_VCPU_WAIT_FOR_INTERRUPT,
+ vcpu_state_blocked_interrupt);
+}
- /* Switch back to primary VM. */
- vm_set_current(&primary_vm);
+/**
+ * Configures the VM to send/receive data through the specified pages. The pages
+ * must not be shared.
+ */
+int32_t api_vm_configure(paddr_t send, paddr_t recv)
+{
+ struct vm *vm = cpu()->current->vm;
+ int32_t ret;
+
+ /* Fail if addresses are not page-aligned. */
+ if ((recv & (PAGE_SIZE - 1)) || (send & (PAGE_SIZE - 1))) {
+ return -1;
+ }
+
+ sl_lock(&vm->lock);
+
+ /* We only allow these to be setup once. */
+ if (vm->rpc.recv || vm->rpc.send) {
+ ret = -1;
+ goto exit;
+ }
/*
- * Inidicate to primary VM that this vcpu blocked waiting for an
- * interrupt.
+ * TODO: Once memory sharing is implemented, we need to make sure that
+ * these pages aren't and won't be shared.
*/
- arch_regs_set_retval(&vcpu->regs, HF_VCPU_WAIT_FOR_INTERRUPT);
- return vcpu;
+ /*
+ * Check that both pages are acessible from the VM, i.e., ensure that
+ * the caller isn't try to use another VM's memory.
+ */
+ if (!mm_ptable_is_mapped(&vm->ptable, recv, 0) ||
+ !mm_ptable_is_mapped(&vm->ptable, send, 0)) {
+ ret = -1;
+ goto exit;
+ }
+
+ /* Map the send page as read-only in the hypervisor address space. */
+ if (!mm_map((vaddr_t)send, (vaddr_t)send + PAGE_SIZE, send,
+ MM_MODE_R)) {
+ ret = -1;
+ goto exit;
+ }
+
+ /*
+ * Map the receive page as writable in the hypervisor address space. On
+ * failure, unmap the send page before returning.
+ */
+ if (!mm_map((vaddr_t)recv, (vaddr_t)recv + PAGE_SIZE, recv,
+ MM_MODE_W)) {
+ mm_unmap((vaddr_t)send, (vaddr_t)send + PAGE_SIZE, 0);
+ ret = -1;
+ goto exit;
+ }
+
+ /* Save pointers to the pages. */
+ vm->rpc.send = (const void *)(vaddr_t)send;
+ vm->rpc.recv = (void *)(vaddr_t)recv;
+
+ /* TODO: Notify any waiters. */
+
+ ret = 0;
+exit:
+ sl_unlock(&vm->lock);
+
+ return ret;
+}
+
+/**
+ * Sends an RPC request from the primary VM to a secondary VM. Data is copied
+ * from the caller's send buffer to the destination's receive buffer.
+ */
+int32_t api_rpc_request(uint32_t vm_idx, size_t size)
+{
+ struct vm *from = cpu()->current->vm;
+ struct vm *to;
+ const void *from_buf;
+ int32_t ret;
+
+ /* Basic argument validation. */
+ if (size > PAGE_SIZE || vm_idx >= secondary_vm_count) {
+ return -1;
+ }
+
+ /* Only the primary VM can make calls. */
+ if (from != &primary_vm) {
+ return -1;
+ }
+
+ /*
+ * Check that the sender has configured its send buffer. It is safe to
+ * use from_buf after releasing the lock because the buffer cannot be
+ * modified once it's configured.
+ */
+ sl_lock(&from->lock);
+ from_buf = from->rpc.send;
+ sl_unlock(&from->lock);
+ if (!from_buf) {
+ return -1;
+ }
+
+ to = secondary_vm + vm_idx;
+ sl_lock(&to->lock);
+
+ if (to->rpc.state != rpc_state_idle || !to->rpc.recv) {
+ /* Fail if the target isn't currently ready to receive data. */
+ ret = -1;
+ } else {
+ /* Copy data. */
+ memcpy(to->rpc.recv, from_buf, size);
+ to->rpc.recv_bytes = size;
+
+ if (!to->rpc.recv_waiter) {
+ to->rpc.state = rpc_state_pending;
+ ret = 0;
+ } else {
+ struct vcpu *to_vcpu = to->rpc.recv_waiter;
+
+ to->rpc.state = rpc_state_inflight;
+
+ /*
+ * Take target vcpu out of waiter list and mark as ready
+ * to run again.
+ */
+ sl_lock(&to_vcpu->lock);
+ to->rpc.recv_waiter = to_vcpu->rpc_next;
+ to_vcpu->state = vcpu_state_ready;
+ arch_regs_set_retval(&to_vcpu->regs, size);
+ sl_unlock(&to_vcpu->lock);
+
+ ret = to_vcpu - to->vcpus + 1;
+ }
+ }
+
+ sl_unlock(&to->lock);
+
+ return ret;
+}
+
+/**
+ * Reads a request sent from a previous call to api_rpc_request. If one isn't
+ * available, this function can optionally block the caller until one becomes
+ * available.
+ *
+ * Once the caller has completed handling a request, it must indicate it by
+ * either calling api_rpc_reply or api_rpc_ack. No new requests can be accepted
+ * until the current one is acknowledged.
+ */
+int32_t api_rpc_read_request(bool block, struct vcpu **next)
+{
+ struct vcpu *vcpu = cpu()->current;
+ struct vm *vm = vcpu->vm;
+ int32_t ret;
+
+ /* Only the secondary VMs can receive calls. */
+ if (vm == &primary_vm) {
+ return -1;
+ }
+
+ sl_lock(&vm->lock);
+ if (vm->rpc.state == rpc_state_pending) {
+ ret = vm->rpc.recv_bytes;
+ vm->rpc.state = rpc_state_inflight;
+ } else if (!block) {
+ ret = -1;
+ } else {
+ sl_lock(&vcpu->lock);
+ vcpu->state = vcpu_state_blocked_rpc;
+
+ /* Push vcpu into waiter list. */
+ vcpu->rpc_next = vm->rpc.recv_waiter;
+ vm->rpc.recv_waiter = vcpu;
+ sl_unlock(&vcpu->lock);
+
+ /* Switch back to primary vm. */
+ *next = &primary_vm.vcpus[cpu_index(cpu())];
+ vm_set_current(&primary_vm);
+
+ /*
+ * Inidicate to primary VM that this vcpu blocked waiting for an
+ * interrupt.
+ */
+ arch_regs_set_retval(&(*next)->regs,
+ HF_VCPU_WAIT_FOR_INTERRUPT);
+ ret = 0;
+ }
+ sl_unlock(&vm->lock);
+
+ return ret;
+}
+
+/**
+ * Sends a reply from a secondary VM to the primary VM. Data is copied from the
+ * caller's send buffer to the destination's receive buffer.
+ *
+ * It can optionally acknowledge the pending request.
+ */
+int32_t api_rpc_reply(size_t size, bool ack, struct vcpu **next)
+{
+ struct vm *from = cpu()->current->vm;
+ struct vm *to;
+ const void *from_buf;
+ /* Basic argument validation. */
+ if (size > PAGE_SIZE) {
+ return -1;
+ }
+
+ /* Only the secondary VM can send responses. */
+ if (from == &primary_vm) {
+ return -1;
+ }
+
+ /* Acknowledge the current pending request if requested. */
+ if (ack) {
+ api_rpc_ack();
+ }
+
+ /*
+ * Check that the sender has configured its send buffer. It is safe to
+ * use from_buf after releasing the lock because the buffer cannot be
+ * modified once it's configured.
+ */
+ sl_lock(&from->lock);
+ from_buf = from->rpc.send;
+ sl_unlock(&from->lock);
+ if (!from_buf) {
+ return -1;
+ }
+
+ to = &primary_vm;
+ sl_lock(&to->lock);
+
+ if (to->rpc.state != rpc_state_idle || !to->rpc.recv) {
+ /*
+ * Fail if the target isn't currently ready to receive a
+ * response.
+ */
+ sl_unlock(&to->lock);
+ return -1;
+ }
+
+ /* Copy data. */
+ memcpy(to->rpc.recv, from_buf, size);
+ to->rpc.recv_bytes = size;
+ to->rpc.state = rpc_state_inflight;
+ sl_unlock(&to->lock);
+
+ /*
+ * Switch back to primary VM so that it is aware that a response was
+ * received. But we leave the current vcpu still runnable.
+ */
+ *next = api_switch_to_primary((size << 8) | HF_VCPU_RESPONSE_READY,
+ vcpu_state_ready);
+
+ return 0;
+}
+
+/**
+ * Acknowledges that either a request or a reply has been received and handled.
+ * After this call completes, the caller will be able to receive additional
+ * requests or replies.
+ */
+int32_t api_rpc_ack(void)
+{
+ struct vm *vm = cpu()->current->vm;
+ int32_t ret;
+
+ sl_lock(&vm->lock);
+ if (vm->rpc.state != rpc_state_inflight) {
+ ret = -1;
+ } else {
+ ret = 0;
+ vm->rpc.state = rpc_state_idle;
+ }
+ sl_unlock(&vm->lock);
+
+ if (ret == 0) {
+ /* TODO: Notify waiters, if any. */
+ }
+
+ return ret;
}
diff --git a/src/arch/aarch64/handler.c b/src/arch/aarch64/handler.c
index 6c00067..3eec780 100644
--- a/src/arch/aarch64/handler.c
+++ b/src/arch/aarch64/handler.c
@@ -38,8 +38,9 @@
}
default:
- dlog("Unknown sync exception pc=0x%x, esr=0x%x, ec=0x%x\n", elr,
- esr, esr >> 26);
+ dlog("Unknown current sync exception pc=0x%x, esr=0x%x, "
+ "ec=0x%x\n",
+ elr, esr, esr >> 26);
for (;;) {
/* do nothing */
}
@@ -79,6 +80,26 @@
ret.user_ret = api_vcpu_run(arg1, arg2, &ret.new);
break;
+ case HF_VM_CONFIGURE:
+ ret.user_ret = api_vm_configure(arg1, arg2);
+ break;
+
+ case HF_RPC_REQUEST:
+ ret.user_ret = api_rpc_request(arg1, arg2);
+ break;
+
+ case HF_RPC_READ_REQUEST:
+ ret.user_ret = api_rpc_read_request(arg1, &ret.new);
+ break;
+
+ case HF_RPC_ACK:
+ ret.user_ret = api_rpc_ack();
+ break;
+
+ case HF_RPC_REPLY:
+ ret.user_ret = api_rpc_reply(arg1, arg2, &ret.new);
+ break;
+
default:
ret.user_ret = -1;
}
@@ -90,10 +111,8 @@
{
/* TODO: Only switch if we know the interrupt was not for the secondary
* VM. */
-
/* Switch back to primary VM, interrupts will be handled there. */
- vm_set_current(&primary_vm);
- return &primary_vm.vcpus[cpu_index(cpu())];
+ return api_switch_to_primary(HF_VCPU_YIELD, vcpu_state_ready);
}
struct vcpu *sync_lower_exception(uint64_t esr)
@@ -124,8 +143,25 @@
/* do nothing */
}
+ case 0x20: /* EC = 100000, Instruction abort. */
+ dlog("Instruction abort: pc=0x%x, esr=0x%x, ec=0x%x",
+ vcpu->regs.pc, esr, esr >> 26);
+ if (!(esr & (1u << 10))) { /* Check FnV bit. */
+ dlog(", far=0x%x, hpfar=0x%x", read_msr(far_el2),
+ read_msr(hpfar_el2) << 8);
+ } else {
+ dlog(", far=invalid");
+ }
+
+ dlog(", vttbr_el2=0x%x", read_msr(vttbr_el2));
+ dlog("\n");
+ for (;;) {
+ /* do nothing */
+ }
+
default:
- dlog("Unknown sync exception pc=0x%x, esr=0x%x, ec=0x%x\n",
+ dlog("Unknown lower sync exception pc=0x%x, esr=0x%x, "
+ "ec=0x%x\n",
vcpu->regs.pc, esr, esr >> 26);
for (;;) {
/* do nothing */
diff --git a/src/arch/aarch64/inc/arch_api.h b/src/arch/aarch64/inc/arch_api.h
index 82b1429..51a9880 100644
--- a/src/arch/aarch64/inc/arch_api.h
+++ b/src/arch/aarch64/inc/arch_api.h
@@ -8,11 +8,17 @@
#define HF_VCPU_YIELD 0x00
#define HF_VCPU_WAIT_FOR_INTERRUPT 0x01
#define HF_VCPU_WAKE_UP 0x02
+#define HF_VCPU_RESPONSE_READY 0x03
/* TODO: Define constants below according to spec. */
-#define HF_VCPU_RUN 0xff00
-#define HF_VM_GET_COUNT 0xff01
-#define HF_VCPU_GET_COUNT 0xff02
+#define HF_VCPU_RUN 0xff00
+#define HF_VM_GET_COUNT 0xff01
+#define HF_VCPU_GET_COUNT 0xff02
+#define HF_VM_CONFIGURE 0xff03
+#define HF_RPC_REQUEST 0xff04
+#define HF_RPC_READ_REQUEST 0xff05
+#define HF_RPC_ACK 0xff06
+#define HF_RPC_REPLY 0xff07
/* clang-format on */
diff --git a/src/cpu.c b/src/cpu.c
index cab40b8..581bc11 100644
--- a/src/cpu.c
+++ b/src/cpu.c
@@ -94,19 +94,20 @@
memset(vcpu, 0, sizeof(*vcpu));
sl_init(&vcpu->lock);
vcpu->vm = vm;
+ vcpu->state = vcpu_state_off;
/* TODO: Initialize vmid register. */
}
void vcpu_on(struct vcpu *vcpu)
{
sl_lock(&vcpu->lock);
- vcpu->is_on = true;
+ vcpu->state = vcpu_state_ready;
sl_unlock(&vcpu->lock);
}
void vcpu_off(struct vcpu *vcpu)
{
sl_lock(&vcpu->lock);
- vcpu->is_on = false;
+ vcpu->state = vcpu_state_off;
sl_unlock(&vcpu->lock);
}
diff --git a/src/mm.c b/src/mm.c
index afdbf50..00c5d3f 100644
--- a/src/mm.c
+++ b/src/mm.c
@@ -341,6 +341,50 @@
}
/**
+ * Determines if the given virtual address is mapped in the given page table
+ * by recursively traversing all levels of the page table.
+ */
+static bool mm_is_mapped_recursive(const pte_t *table, vaddr_t addr, int level)
+{
+ pte_t pte;
+ vaddr_t va_level_end = mm_level_end(addr, level);
+
+ /* It isn't mapped if it doesn't fit in the table. */
+ if (addr >= va_level_end) {
+ return false;
+ }
+
+ pte = table[mm_index(addr, level)];
+
+ if (level == 0) {
+ return arch_mm_pte_is_present(pte);
+ }
+
+ if (arch_mm_is_block_allowed(level) && arch_mm_pte_is_block(pte)) {
+ return true;
+ }
+
+ if (arch_mm_pte_is_table(pte)) {
+ return mm_is_mapped_recursive(arch_mm_pte_to_table(pte), addr,
+ level - 1);
+ }
+
+ return false;
+}
+
+/**
+ * Determines if the given virtual address is mapped in the given page table.
+ */
+bool mm_ptable_is_mapped(struct mm_ptable *t, vaddr_t addr, int mode)
+{
+ int level = arch_mm_max_level(mode);
+
+ addr = arch_mm_clear_va(addr);
+
+ return mm_is_mapped_recursive(t->table, addr, level);
+}
+
+/**
* Initialises the given page table.
*/
bool mm_ptable_init(struct mm_ptable *t, uint32_t id, int mode)
diff --git a/src/vm.c b/src/vm.c
index b1d26c9..cadda86 100644
--- a/src/vm.c
+++ b/src/vm.c
@@ -1,12 +1,16 @@
#include "vm.h"
#include "cpu.h"
+#include "std.h"
bool vm_init(struct vm *vm, uint32_t id, uint32_t vcpu_count)
{
uint32_t i;
+ memset(vm, 0, sizeof(*vm));
+
vm->vcpu_count = vcpu_count;
+ vm->rpc.state = rpc_state_idle;
/* Do basic initialization of vcpus. */
for (i = 0; i < vcpu_count; i++) {